# Operant — full content > Custom Data Collection for Robotics and Physical AI. Operant designs and runs custom real-world data collection for robotics and physical AI teams, multi-sensor capture, teleoperation, and edge-case scenarios built to your spec. Operant is not a dataset marketplace, catalog, or data broker. This file contains the full text of every page for answer engines and LLM retrieval. Source of truth: https://www.operantdata.com/. --- # Services ## Edge-Case Data Collection for Robotics URL: https://www.operantdata.com/services/edge-case-scenarios Capture rare failures, near-misses, recoveries, and long-tail behaviors so evaluation reflects operational risk, not just clean demos. Edge-case data collection is the deliberate capture of rare failures, near-misses, recoveries, and long-tail behaviors so evaluation reflects real operational risk rather than clean demonstrations. Operant co-defines a failure taxonomy and sampling targets with your team, then runs scripted and opportunistic capture under safety review. The result is tail-event libraries that make your evaluation honest about how a policy behaves when things go wrong. ## Why tail events matter A policy that succeeds on 95% of happy-path episodes can still be unsafe, because deployment risk concentrates in the rare 5%: slips, collisions, dropped objects, planner stalls, and the recoveries that follow. Evaluation that ignores the tail overstates reliability. Capturing tail events deliberately is the only way to measure it. ## Failure taxonomy We start by co-defining a taxonomy of the failures that matter for your system, then map each to capture targets. This shared vocabulary keeps collection, labeling, and evaluation aligned across the program. ## Sampling ratios Because failures are rare, we set explicit sampling targets and combine scripted scenarios with opportunistic capture to hit them. This pairs naturally with [sim-to-real data collection](/services/sim-to-real-gap), since many tail events are exactly what simulation fails to reproduce. ## Safety protocols Capturing failures safely requires protocols agreed before any capture begins, review of risky scenarios, controlled environments, and clear stop conditions. Safety is scoped alongside the taxonomy. ## Delivery and eval usage Tail events are labeled and organized into evaluation slices, delivered with metadata and provenance. See [warehouse defective-SKU pick failures](/datasets/defective-sku-pick-failures) for a concrete failure-capture scenario, and [eval benchmarks vs. the real world](/blog/eval-benchmarks-real-world) for how tail data sharpens evaluation. Edge-case capture is a core part of any serious [robotics data collection](/services/robotics-data-collection) program. --- ## Imitation Learning Data Collection URL: https://www.operantdata.com/services/imitation-learning-data-collection Scope, capture, and QA robot demonstrations for imitation learning, with guidance on teleoperation, episode boundaries, labels, and evaluation. Imitation learning data collection is the work of capturing expert robot demonstrations, observations paired with the actions an expert took, so a policy can learn to reproduce them. Done well, it requires demonstrations in your robot's action space, consistent episode boundaries, synchronized sensors, and metadata you can filter on. Operant scopes, captures, and QAs these programs end to end, primarily through teleoperation and human demonstration. ## IL data requirements Imitation learning is only as good as its demonstrations. The data must match your action space, carry consistent episode boundaries, and include enough operator and scene diversity to generalize. Sensor streams have to be time-aligned so observations and actions correspond exactly. ## Demonstration modalities Most imitation learning data comes from [teleoperation capture](/services/teleoperation-capture), where a human guides the robot, or from egocentric human demonstration. We record synchronized video, depth, proprioception, and control signals, with calibration handled through our [multi-sensor synchronization service](/services/multi-sensor-sync). ## Episode design Episode boundaries, reset conditions, and success criteria need to be defined before capture, not after. Inconsistent episodes are one of the most common and expensive mistakes; we lock these during scoping and validate them in the pilot. ## Metadata and labels Each demonstration is tagged with task, operator, scene, and outcome so you can filter and balance your dataset. Labels are scoped to your schema and applied consistently across the program. ## Common quality failures Action-space mismatch, unsynchronized sensors, low operator diversity, and missing metadata silently degrade policies. See [teleoperation best practices](/blog/teleoperation-best-practices) for how small pilot mistakes become expensive at scale, and [robot demonstration data](/services/robot-demonstration-data) for what a good demonstration package contains. --- ## Multimodal Robotics Data Collection URL: https://www.operantdata.com/services/multi-sensor-sync Capture tightly synchronized RGB-D, LiDAR, IMU, force/torque, and proprioception data with documented calibration and drift controls. Multimodal robotics data collection is the capture of multiple, tightly time-aligned sensor streams, RGB-D, LiDAR, IMU, force/torque, and proprioception, with documented calibration so observations and actions correspond exactly. Operant designs and validates the sensor rig, characterizes synchronization, and runs automated drift checks during scaled capture. The result is multimodal data your ML engineers can fuse and trust, not streams that silently drift out of alignment. ## Why synchronization matters When sensor streams are not synchronized, the link between what the robot saw and what it did is corrupted, and models learn from misaligned data. For manipulation, sensor fusion, and any timing-dependent policy, sub-millisecond alignment is the difference between usable and wasted data. Synchronization is a first-class deliverable in every Operant program. ## Supported modalities We capture RGB-D camera arrays, LiDAR, IMU, force/torque, audio, and proprioceptive and control signals, aligned to your exact hardware list. This service underpins [teleoperation capture](/services/teleoperation-capture) and broader [robotics data collection](/services/robotics-data-collection) programs. ## Calibration and drift checks Each rig is calibrated, intrinsics and extrinsics, and synchronization is characterized during a pilot. During scaled collection we run automated drift checks and report against agreed tolerances, so problems surface immediately rather than in your training logs weeks later. ## File formats and metadata Deliverables include synchronized logs, calibration files, and scene-level metadata in the formats your stack expects. Episodes carry the identifiers needed to filter, balance, and audit the dataset. ## Common failure modes The usual culprits, unflagged clock drift, uncalibrated extrinsics, dropped frames, and incomplete metadata, are exactly what our QA gates catch. For programs targeting deployment gaps, pair this with [sim-to-real data collection](/services/sim-to-real-gap); for a concrete multimodal scenario, see [AV rain LiDAR and camera capture](/datasets/av-rain-lidar-camera). --- ## Physical AI Data Collection URL: https://www.operantdata.com/services/physical-ai-data-collection Real-world capture programs for physical AI, robots, sensors, environments, and evaluation slices that match deployment, not internet-scale proxies. Physical AI data collection is the work of capturing real-world sensor, action, and outcome data so embodied systems learn and are evaluated against the conditions they will actually face. Operant designs capture programs for robots, sensors, and environments, teleoperation, multimodal logs, egocentric views, and failure data, matched to your deployment. The result is training and evaluation data your models can trust, not internet-scale proxies that miss your embodiment. ## What physical AI data includes Physical AI spans manipulation, locomotion, navigation, and human interaction. The data that supports it includes teleoperation and demonstration trajectories, synchronized RGB-D, LiDAR, IMU, and force/torque streams, proprioceptive and control signals, and explicitly captured tail events. Each is tied to scene-level metadata so episodes are auditable and reproducible. ## Why internet-scale data is not enough Web-scale video and simulation are powerful for pretraining, but they lack calibrated sensors, action labels, and the embodiment-specific dynamics of contact and timing. A policy trained only on those proxies tends to break on the [sim-to-real gap](/services/sim-to-real-gap): lighting, wear, contact, and human interference that simulation does not reproduce. Targeted real-world capture closes that gap where it actually lives. ## Teleop, egocentric, multimodal, and failure data - **Teleoperation:** human-guided demonstrations for imitation learning, via our [teleoperation capture service](/services/teleoperation-capture). - **Egocentric:** first-person capture of target behaviors in real environments. - **Multimodal:** tightly synchronized sensor suites with calibration and drift control. - **Failure data:** rare events, near-misses, and recoveries that dominate deployment risk. ## Data quality checklist A defensible physical AI program defines, up front: time-sync tolerances, calibration procedures, metadata schema, diversity and coverage targets, and acceptance criteria. Operant agrees these during scoping and reports against them at handoff. ## Sample program design A representative program scopes target behaviors and environments, runs a two-to-four week pilot to validate sync and metadata, then scales capture with QA checkpoints. See how this maps to verticals like [humanoid robotics](/industries/humanoid-robotics) and broader [robotics data collection](/services/robotics-data-collection). --- ## Robot Demonstration Data URL: https://www.operantdata.com/services/robot-demonstration-data Capture demonstration data matched to your robot, environment, and evaluation goals, not generic datasets with mismatched action spaces. Robot demonstration data is example episodes, observations paired with the actions taken, captured so a policy can learn your task. The data that works is matched to your robot, your environment, and your evaluation goals, not a generic dataset with a mismatched action space. Operant captures human, teleoperation, and egocentric demonstrations with clear episode boundaries, synchronized sensors, and audit-ready metadata. ## What counts as a good demonstration A good demonstration is recorded in your robot's action space, has clear start and end boundaries, is synchronized across sensors, and carries metadata describing task, operator, scene, and outcome. Demonstrations that miss any of these silently undermine downstream training. ## Human vs. teleop vs. egocentric - **Human demonstration:** a person performs the task; useful for reference and egocentric pipelines. - **Teleoperation:** an operator drives the robot directly, captured via our [teleoperation capture service](/services/teleoperation-capture). - **Egocentric:** first-person capture of the behavior in a real environment. We scope the right mix for your policy as part of [imitation learning data collection](/services/imitation-learning-data-collection). ## Annotation options Demonstrations can be delivered raw or with labels scoped to your schema, segmentation, phase tags, success annotations, or object references, applied consistently across the program. ## Delivery formats You receive synchronized sensor logs, calibration files, episode metadata, and optional labels in the formats your training stack expects, with documentation and provenance. For a concrete scenario, see [dual-arm kitchen manipulation](/datasets/dual-arm-kitchen-manipulation) or the broader [robotics training data](/services/robotics-training-data) guide. --- ## Robotics Data Collection for Physical AI URL: https://www.operantdata.com/services/robotics-data-collection Custom robotics data collection programs, teleoperation, synchronized sensors, edge cases, QA, and pilot-to-production delivery built to your spec. Robotics data collection is the practice of capturing real-world sensor, action, and outcome data so physical AI models learn and are evaluated against the conditions they actually face. Operant designs and runs custom collection programs, teleoperation, synchronized multi-sensor capture, and edge-case scenarios, built around your robot, your environments, and your evaluation goals. We move you from pilot to production with documented provenance and QA, never a generic catalog download. ## Why real-world robotics data is scarce Most robotics teams can simulate cleanly and pretrain on web-scale video, yet still fail on deployment. The gap is real-world data that matches your embodiment: contact-rich manipulation, calibrated multi-camera views, proprioception, and the long tail of failures your evaluation has to trust. That data is expensive to capture well, hard to synchronize, and almost never available off the shelf in a form that matches your action space. This is why custom collection exists. Rather than reshaping your problem to fit an existing dataset, a collection program captures exactly the trajectories, sensors, and scenarios your model needs. ## Collection methods Operant supports the methods that map to how physical AI teams actually train and evaluate: - **[Teleoperation and demonstration capture](/services/teleoperation-capture)** for imitation learning and policy fine-tuning. - **Egocentric and human-demonstration capture** where a human performs the target behavior. - **On-robot autonomous logging** during scripted or policy-driven runs. - **Edge-case and failure capture** for the rare events that dominate deployment risk. Methods are combined per program. A manipulation policy might mix teleoperation for the core skill with targeted [edge-case scenarios](/services/edge-case-scenarios) for recovery behaviors. ## Sensor modalities We capture and time-align the modalities your stack consumes, including RGB-D arrays, LiDAR, IMU, force/torque, audio, and proprioceptive and control streams. Synchronization and calibration are first-class deliverables, not afterthoughts, handled through our [multi-sensor synchronization service](/services/multi-sensor-sync) with documented drift characterization. ## QA and provenance Every program ships with quality gates agreed during scoping: time-sync tolerances, calibration checks, metadata completeness, and diversity targets. You receive QA reports, calibration files, and scene-level metadata so your ML team can audit and reproduce what was collected. ## Pilot-to-production workflow 1. **Scope** environments, sensors, behaviors, volume, and acceptance criteria. 2. **Pilot** a short capture to validate calibration, labeling, and integration. 3. **Scale** production collection with QA checkpoints and edge-case coverage. 4. **Handoff** deliverables in your formats with documentation and provenance. ## Industry use cases Robotics data collection looks different across verticals. See how it applies to [humanoid robotics](/industries/humanoid-robotics), [warehouse automation](/industries/warehouse-automation), and [autonomous vehicles](/industries/autonomous-vehicles), or browse capture scenarios such as [warehouse pallet pick teleoperation](/datasets/warehouse-pallet-pick-teleoperation). --- ## Robotics Training Data URL: https://www.operantdata.com/services/robotics-training-data A buyer-focused guide to robotics training data, demonstrations, teleoperation, synchronization, labels, evaluation slices, and deployment-fit collection. Robotics training data is the structured sensor and action data used to train and evaluate robot policies: teleoperation trajectories, human demonstrations, synchronized multimodal logs, and the metadata that makes episodes usable. The data that actually moves real-world performance is matched to your robot, calibrated, and rich in the tail events your evaluation must reflect. This guide covers dataset types, collection versus annotation, what good looks like, and how to choose a capture partner. ## Dataset types Robotics training data spans several forms: teleoperation and demonstration datasets for imitation learning, autonomous on-robot logs, multimodal sensor datasets, and dedicated evaluation sets built around tail behaviors. Each serves a different stage, from pretraining representations to validating a policy before deployment. ## Collection vs. annotation Collection produces the raw episodes from the real world; annotation adds labels, segmentation, and structure. Teams often conflate the two and end up with mislabeled or mismatched data. Operant focuses on custom [robotics data collection](/services/robotics-data-collection) and scopes annotation to your label schema so the two stay aligned. ## What "good" data looks like Good robotics training data is matched to your embodiment and action space, time-synchronized and calibrated through a process like our [multi-sensor synchronization service](/services/multi-sensor-sync), complete in metadata, and deliberately diverse. It includes the rare events that determine deployment risk rather than only clean demonstrations. ## Benchmark and eval design Training data is only half the problem. Evaluation slices, curated subsets that stress specific conditions, determine whether you can trust a policy. We design eval slices alongside collection; see [eval benchmarks vs. the real world](/blog/eval-benchmarks-real-world) for the rationale. ## Vendor selection checklist When choosing a robotics training data partner, confirm: embodiment and action-space fit, calibration and sync rigor, metadata and provenance, edge-case coverage, data ownership terms, and a pilot-before-scale workflow. Operant is built around each of these. To go deeper on demonstrations, see [imitation learning data collection](/services/imitation-learning-data-collection). --- ## Sim-to-Real Data Collection for Robotics URL: https://www.operantdata.com/services/sim-to-real-gap Close simulation gaps with targeted real-world capture in lighting, contact, wear, and human-interference conditions that break policies after deployment. Sim-to-real data collection closes the gap between simulation and the real world by capturing targeted real episodes in exactly the conditions simulation gets wrong, lighting, contact physics, wear, sensor noise, and human interference. Operant runs a gap analysis, prioritizes the highest-risk domains, and ties each capture target to a measurable evaluation slice. The goal is a policy that holds up after deployment, with improvement you can actually measure. ## Where sim breaks Strong simulation pipelines still miss the long tail: subtle contact dynamics, lighting and material variation, sensor noise, wear over time, and unpredictable human behavior. These are where deployed policies fail, and where targeted real-world capture pays off. Simulation tells you what should happen; real capture tells you what does. ## Gap analysis framework We start with a gap analysis workshop to identify where simulation coverage is thin or misleading, then produce a prioritized scenario list ranked by deployment risk. This keeps capture focused on the domains that move your metrics rather than collecting broadly and hoping. ## Real-world slice design Each prioritized gap becomes a capture slice with defined conditions, sensors, and acceptance criteria. Multimodal capture is handled through our [multi-sensor synchronization service](/services/multi-sensor-sync), and rare events through [edge-case data collection](/services/edge-case-scenarios). ## Metrics to track We define evaluation slices before capture and track policy performance on real held-out data before and after, so the value of new data is measurable. See [sim-to-real metrics](/blog/sim-to-real-metrics) for how we frame this. ## Example programs A typical program pilots capture in the single highest-risk domain, validates the eval improvement, then scales across the prioritized list. This fits naturally into broader [robotics data collection](/services/robotics-data-collection) and applies directly to verticals like [industrial manipulation](/industries/industrial-manipulation). --- ## Teleoperation Capture URL: https://www.operantdata.com/services/teleoperation-capture Human-guided teleoperation and demonstration logging for imitation learning and policy fine-tuning, with multi-camera sync, metadata, and QA gates. Teleoperation capture is the collection of human-guided robot demonstrations, an operator drives the robot while synchronized sensors record observations, actions, and outcomes, for imitation learning and policy fine-tuning. Operant runs teleoperation programs around your action space, camera setup, and quality bar, with operator diversity, calibration, and QA gates agreed up front. You get demonstrations matched to your robot, not generic trajectories with a mismatched action space. ## When teleop is the right method Teleoperation is the right method when you need expert demonstrations in your robot's action space, particularly for contact-rich manipulation, dexterous tasks, or behaviors that are difficult to script autonomously. It is the workhorse of [imitation learning data collection](/services/imitation-learning-data-collection), and it pairs well with targeted [edge-case capture](/services/edge-case-scenarios) for recovery behaviors. ## Action-space design The action space, what the operator controls and how it maps to the robot, is the most consequential decision in a teleoperation program. We align on it during scoping so demonstrations transfer cleanly to training. Mismatched action spaces are a leading cause of policies that look fine in training and fail on hardware. ## Camera and sensor setup We design and calibrate the camera rig, typically multi-camera RGB-D, alongside proprioception and control streams, with extrinsic and intrinsic calibration captured as deliverables. Time synchronization is handled through our [multi-sensor synchronization service](/services/multi-sensor-sync) so observations and actions correspond exactly. ## QA and operator diversity Quality comes from process. We set time-sync tolerances, metadata completeness checks, and operator diversity targets, then track operator identity and scene variation so the dataset generalizes. A calibration pilot validates the full pipeline before scaling. ## Deliverables You receive synchronized video, depth, proprioception, and control logs, calibration files, episode metadata and scene catalogs, and optional labels, in the formats your pipeline expects, with documentation and provenance. ## Where teleoperation fits Teleoperation programs power [humanoid robotics](/industries/humanoid-robotics) and [warehouse automation](/industries/warehouse-automation) capture alike. For practical guidance on avoiding expensive pilot mistakes, see [teleoperation best practices](/blog/teleoperation-best-practices). --- # Industries ## Autonomous Vehicle Data Collection URL: https://www.operantdata.com/industries/autonomous-vehicles On-road and closed-course multimodal AV capture for lighting tails, weather, interventions, and geospatially documented evaluation slices. Autonomous vehicle data collection captures the on-road and closed-course driving data AV stacks need, multimodal sensor logs, weather and lighting tails, and intervention episodes, with sensor parity to production vehicles and rigorous geospatial and consent metadata. Operant scopes capture around your sensor rig and evaluation goals, with documented consent for public-road work where applicable. The result is AV data that stresses the conditions where policies actually fail. ## Sensor parity and ground truth AV data is only useful if the capture rig matches the production vehicle's sensors. We design for parity, calibrated multi-camera, LiDAR, IMU, and radar where relevant, time-aligned through our [multi-sensor synchronization service](/services/multi-sensor-sync), with ground-truth and geospatial metadata. ## Weather and light tails Rain, fog, glare, and low light are where perception degrades. We capture these tails deliberately, illustrated by [AV rain LiDAR and camera capture](/datasets/av-rain-lidar-camera), so evaluation reflects adverse conditions rather than clear-day driving. ## Intervention and safety capture Disengagements and interventions are high-value, rare episodes. We capture them through [edge-case data collection](/services/edge-case-scenarios) and label them for policy evaluation, under safety protocols agreed up front. ## Consent and compliance Public-road capture involves people and places. We follow your policies and applicable regulations, with consent workflows and data minimization, see our [privacy and consent on-site](/blog/privacy-consent-on-site) approach, agreed before capture starts. ## FAQ **Do you match our production sensor rig?** Yes, sensor parity is a core scoping requirement. **Can you capture adverse weather and interventions?** Yes, these tail conditions are explicit capture targets. This vertical is part of Operant's broader [robotics data collection](/services/robotics-data-collection) practice. --- ## Humanoid Robot Data Collection URL: https://www.operantdata.com/industries/humanoid-robotics Whole-body teleoperation, manipulation, locomotion, and recovery capture for humanoid training pipelines in homes, labs, and industrial settings. Humanoid robot data collection captures the whole-body behaviors a humanoid needs to learn, teleoperated manipulation, locomotion, balance recovery, and the tail events that determine safety, in homes, labs, and light industrial settings. Operant scopes capture to your platform's morphology and action space, with whole-body synchronization and fall and recovery episodes. The result is humanoid training data matched to how your robot actually moves and where it will be deployed. ## Key humanoid modalities Humanoids combine locomotion, manipulation, and human interaction, so the data spans more modalities than most platforms. We capture synchronized multi-camera RGB-D, proprioception across many joints, IMU, and force/torque, aligned through our [multi-sensor synchronization service](/services/multi-sensor-sync) so whole-body state and action correspond exactly. ## Locomotion and manipulation capture We capture locomotion on varied flooring and terrain alongside reach-and-grasp and bimanual tasks in clutter, with force-relevant metadata. Demonstrations are most often gathered through [teleoperation capture](/services/teleoperation-capture), the workhorse method for dexterous, contact-rich humanoid behaviors. Concrete scenarios include [humanoid stair ascent](/datasets/humanoid-stair-ascend) and [humanoid home navigation](/datasets/humanoid-home-navigation). ## Recovery and tail events Humanoids fall, slip, and need to recover. Deliberately capturing these tail events through [edge-case data collection](/services/edge-case-scenarios) is what lets your evaluation reflect real safety risk rather than only successful demonstrations. ## Evaluation slices We design evaluation slices alongside collection, terrain types, clutter levels, recovery scenarios, so you can measure policy behavior under the conditions that matter, not just average-case success. ## FAQ **What environments do you capture humanoid data in?** Homes, labs, and light industrial settings, scoped to where your humanoid will operate. **Can you capture falls and recoveries safely?** Yes, under safety protocols agreed during scoping, combining scripted and opportunistic capture. Ready to scope a program? This is part of Operant's broader [robotics data collection](/services/robotics-data-collection) practice. --- ## Industrial Manipulation Data Collection URL: https://www.operantdata.com/industries/industrial-manipulation Assembly, insertion, and material-handling capture in factory cells, with force/torque traces, vision occlusions, and tool-change episodes. Industrial manipulation data collection captures the contact-rich assembly, insertion, and material-handling behaviors robots perform in factory cells, aligned to cycle times, fixturing, and safety interlocks. Operant captures force/torque traces, vision-occluded grasps, and tool-change episodes with the metadata your continuous-improvement loops need. The result is data matched to your cell, your tooling, and the contact dynamics that simulation struggles to reproduce. ## Industrial tasks Factory manipulation spans precise insertion, assembly, fastening, and material handling, often contact-rich and force-dependent. We scope capture to the tasks in your cells, with examples like [industrial peg insertion with force sensing](/datasets/industrial-peg-insertion-force). ## Modalities Industrial tasks depend on force and timing, so force/torque traces and proprioception are captured alongside synchronized vision, aligned through our [multi-sensor synchronization service](/services/multi-sensor-sync). Time alignment is essential when contact events are milliseconds long. ## Contact and occlusion Vision-occluded grasps and contact-rich inserts are where many policies fail. These are exactly the conditions simulation gets wrong, so industrial programs pair naturally with [sim-to-real data collection](/services/sim-to-real-gap) and deliberate [edge-case capture](/services/edge-case-scenarios) of jams, misfeeds, and tool failures. ## Safety and cycle times Capture aligns to cycle times, fixturing, and safety interlocks so collection fits production constraints rather than disrupting them. ## FAQ **Do you capture force/torque data?** Yes, force and torque traces are central to industrial manipulation capture. **Can capture fit our cycle times and interlocks?** Yes, capture is scoped around your cell's constraints during planning. This vertical is part of Operant's broader [robotics data collection](/services/robotics-data-collection) practice. --- ## Warehouse Robotics Data Collection URL: https://www.operantdata.com/industries/warehouse-automation Capture teleoperation, pick-place, AMR, pallet, and conveyor data in real warehouse settings with operational variability and synchronized sensors. Warehouse robotics data collection captures the picking, placing, palletizing, and AMR navigation behaviors that robots perform in real distribution environments, with the SKU, lighting, and aisle variability that operations actually contain. Operant runs capture during live operations or staged lanes, with synchronized sensors and deliberate tail-event coverage. The result is warehouse training data that reflects operational reality, not a sanitized lab. ## Warehouse tasks Warehouse automation spans manipulation and navigation: grasping varied SKUs, tote and bin picking, conveyor handoffs, pallet building, and autonomous mobile robot routing through shared aisles. We scope capture across the tasks your fleet performs, with examples like [bin picking teleoperation capture](/datasets/bin-picking-teleoperation-capture), [warehouse pallet pick teleoperation](/datasets/warehouse-pallet-pick-teleoperation), and [conveyor handoff](/datasets/warehouse-conveyor-handoff). ## Modalities We capture synchronized multi-camera RGB-D, depth, proprioception, and control streams, plus AMR sensor stacks, aligned through our [multi-sensor synchronization service](/services/multi-sensor-sync). Most manipulation demonstrations are gathered via [teleoperation capture](/services/teleoperation-capture). ## Tail behaviors Warehouses are full of edge cases, SKU variability, dropped items, occluded grasps, and aisle interference. We capture these deliberately through [edge-case data collection](/services/edge-case-scenarios) so your evaluation reflects the failures that actually slow throughput. AMR obstacle handling is illustrated by [AMR aisle obstacle navigation](/datasets/amr-aisle-obstacle-navigation). ## Safety and ops Capturing during live operations requires safety review and lane scheduling, agreed during scoping, so collection never disrupts throughput or safety. ## FAQ **Can capture run during live operations?** Yes, with safety review and lane scheduling agreed during scoping; staged lanes are also an option. **Do you handle SKU and lighting variability?** Yes, operational variability is a capture target, not something to be controlled away. This vertical is part of Operant's broader [robotics data collection](/services/robotics-data-collection) practice. --- # Data scenarios ## AMR aisle navigation with dynamic obstacles URL: https://www.operantdata.com/datasets/amr-aisle-obstacle-navigation AMR navigation with dynamic obstacles in warehouse aisles. Fleet-relevant AMR runs with human and equipment interference in warehouse aisles, captured with intervention and replan metadata. This is a custom capture program for autonomous mobile robot (AMR) navigation through warehouse aisles full of dynamic obstacles. We record LiDAR and RGB-D runs with the human and equipment interference that triggers stops, slowdowns, and replans, tagged with intervention and geospatial metadata. It targets the messy, shared-space navigation that simulation under-represents and that drives most real-world disengagements. ## What we collect Dynamic obstacle fields, stop/start interventions, and replan episodes with geospatial metadata, capturing the interference patterns, people, forklifts, spills, parked equipment, that real aisles contain. ## Sensors and modalities LiDAR and RGB-D matched to your fleet's sensor stack, time-aligned through our [multi-sensor synchronization service](/services/multi-sensor-sync) so perception and motion correspond. ## How capture works A pilot confirms rig parity and intervention labeling, then capture scales across aisles, shifts, and traffic conditions following our [robotics data collection](/services/robotics-data-collection) workflow. ## QA and metadata Episodes carry obstacle type, intervention and replan tags, aisle IDs, and geospatial metadata. QA validates sensor sync, calibration, and label completeness against agreed criteria. ## Who it is for AMR and fleet teams working on [warehouse automation](/industries/warehouse-automation) where navigation must hold up in shared, dynamic spaces. The interference focus pairs with deliberate [edge-case data collection](/services/edge-case-scenarios) for rare near-miss events. --- ## Autonomous driving rain and low-light LiDAR-camera URL: https://www.operantdata.com/datasets/av-rain-lidar-camera Rain and low-light LiDAR-camera driving logs. On-road or closed-course AV capture emphasizing weather and lighting tails, with intervention and disengagement metadata. This is a custom capture program for autonomous driving in rain and low light, the conditions where perception degrades most. We record synchronized LiDAR and camera segments on road or closed course, tagged with intervention and disengagement metadata. Clear-day data overstates perception reliability; these weather and lighting tails are what robust AV evaluation actually needs. ## What we collect Weather-tail segments with intervention and disengagement metadata for perception robustness evals, spanning rain intensity, glare, dusk, and night conditions. ## Sensors and modalities LiDAR and camera matched to your production rig, time-aligned through our [multi-sensor synchronization service](/services/multi-sensor-sync) so fused perception is trustworthy under degraded conditions. ## How capture works A pilot confirms sensor parity, authorization, and labeling, then capture scales across weather windows and routes following our [robotics data collection](/services/robotics-data-collection) workflow with safety drivers. ## QA and metadata Episodes carry weather, lighting, geolocation, and intervention/disengagement tags. QA covers sensor sync, calibration, and label completeness against agreed criteria. ## Who it is for [Autonomous vehicle](/industries/autonomous-vehicles) teams hardening perception against adverse conditions. The tail focus relies on deliberate [edge-case data collection](/services/edge-case-scenarios) for rare interventions. --- ## Bin Picking Teleoperation Capture URL: https://www.operantdata.com/datasets/bin-picking-teleoperation-capture Teleoperated bin picking from cluttered totes and mixed SKU bins. Custom teleoperated bin picking capture for warehouse robots handling cluttered totes, occluded SKUs, and grasp recovery. Bin picking teleoperation capture is a custom data collection program for robots that must pick from cluttered totes, mixed SKU bins, and partially occluded items. Operant records human-guided robot demonstrations with synchronized RGB-D, proprioception, control signals, grasp outcomes, and recovery metadata. The goal is deployment-matched training data for your robot and warehouse conditions, not a reusable marketplace dataset. ## What we collect We collect teleoperated pick attempts from bins, totes, gaylords, and staged warehouse cells that match your deployment constraints. Programs can vary SKU class, packaging material, tote depth, fill level, item orientation, lighting, and occlusion so policies see the clutter that appears during real operations. Each episode can include successful picks, blocked approaches, slips, partial lifts, drops, regrasps, and recovery attempts. Those failure and recovery moments matter for imitation learning because they teach the policy how operators correct course when perception or grasp planning is uncertain. ## Sensors and modalities Typical capture includes multi-camera RGB-D views, wrist or scene cameras, robot proprioception, gripper state, and control streams. Timing and calibration are handled through Operant's [multi-sensor synchronization service](/services/multi-sensor-sync), so observation frames, operator commands, and grasp outcomes align cleanly for training and analysis. If your stack uses a specific sensor placement, action space, gripper, or logging format, the capture plan is built around that interface. The deliverable is a set of trajectories and metadata your ML team can inspect, filter, and feed into its own pipeline. ## How capture works A pilot first validates the camera rig, action-space mapping, episode boundaries, and metadata schema. Operators then capture demonstrations through the [teleoperation capture service](/services/teleoperation-capture), with QA checks for sync tolerances, calibration drift, operator diversity, and metadata completeness. Scaling focuses on variation, not raw repetition. We set collection targets for SKU families, tote configurations, clutter levels, and outcome classes so the dataset covers the cases your policy must generalize across. If recovery behavior is important, failed grasps are sampled deliberately instead of treated as noise. ## QA and metadata Every episode can be tagged with SKU class, tote geometry, fill level, occlusion level, grasp point, grasp outcome, recovery action, operator ID, and scene notes. QA gates compare the delivered files against the statement of work: required modalities, timestamp alignment, calibration files, metadata completeness, and accepted outcome labels. This makes the capture useful for both training and evaluation. Your team can isolate clean demonstrations, study failure clusters, or build held-out test sets around blocked approaches, deformable packaging, or hard-to-see objects. ## Who it is for This scenario is for robotics teams building warehouse pick policies where simulation and open examples do not match the real bin, tote, SKU, and gripper combination. It fits broader [warehouse robotics data collection](/industries/warehouse-automation) programs and complements [warehouse pallet pick teleoperation](/datasets/warehouse-pallet-pick-teleoperation) when a workflow spans both pallet faces and tote-level picking. To scope a bin-picking capture program around your robot, sensors, and SKU mix, [book a discovery call](/book). --- ## Defective SKU pick failure library URL: https://www.operantdata.com/datasets/defective-sku-pick-failures Pick failures on defective or unusual packaging. Structured failure sets for odd-shaped or damaged packaging picks, with a labeled failure taxonomy for evaluation. This is a custom capture program that deliberately builds a library of pick failures on defective or unusual packaging. We capture near-miss grasps, drops, crushes, and recovery attempts, each tagged with the packaging defect that caused it, so your evaluation reflects operational risk. Because failures are rare, they have to be captured on purpose; a happy-path dataset hides exactly the behavior that slows real throughput. ## What we collect Near-miss grasps, drops, and recovery attempts with packaging defect tags, organized into a structured failure taxonomy rather than scattered through a success-heavy log. ## Sensors and modalities Multi-camera RGB-D and proprioception, time-aligned through our [multi-sensor synchronization service](/services/multi-sensor-sync) so the moment of failure is captured precisely. ## How capture works We co-define the failure taxonomy and sampling ratios, then run scripted and opportunistic capture under safety review, following our [edge-case data collection](/services/edge-case-scenarios) methodology. ## QA and metadata Each episode is labeled with failure class, packaging defect, and outcome. Sampling targets and QA gates are set in the statement of work to hit the success-to-failure balance your evaluation needs. ## Who it is for Warehouse teams and evaluation groups building honest benchmarks for [warehouse automation](/industries/warehouse-automation). Complements clean [warehouse pallet pick teleoperation](/datasets/warehouse-pallet-pick-teleoperation) capture by supplying the tail your success data omits. --- ## Dual-arm kitchen manipulation in homes URL: https://www.operantdata.com/datasets/dual-arm-kitchen-manipulation Dual-arm kitchen manipulation in residential environments. Bimanual kitchen task capture in residential layouts with contact-rich interaction, calibrated multi-view video, and object labels. This is a custom capture program for dual-arm kitchen manipulation in real residential homes. We record bimanual cabinet, drawer, and utensil tasks with synchronized RGB-D, force-torque, and proprioception, plus object interaction labels. Real kitchens supply the layout, lighting, and object variety that lab benches cannot, exactly the variation that decides whether domestic manipulation policies generalize. ## What we collect Opening cabinets, transferring utensils, and contact-rich manipulations with calibrated multi-view video, captured bimanually so both arms and their coordination are represented. ## Sensors and modalities Synchronized RGB-D, force-torque, and proprioception, time-aligned through our [multi-sensor synchronization service](/services/multi-sensor-sync), so contact events and motion correspond precisely. ## How capture works Demonstrations are gathered through [teleoperation capture](/services/teleoperation-capture). A pilot validates rig, consent, and labeling, then capture scales across homes and tasks. ## QA and metadata Episodes carry task type, object set, contact outcome, and scene tags. QA covers sync, calibration, and label completeness against agreed criteria. ## Who it is for Home robotics labs and bimanual manipulation teams, adjacent to [humanoid robotics](/industries/humanoid-robotics) work. Provides high-quality [robot demonstration data](/services/robot-demonstration-data) for imitation-learning pipelines. --- ## Humanoid home navigation and reach URL: https://www.operantdata.com/datasets/humanoid-home-navigation Humanoid navigation and reach in furnished homes. Whole-body navigation and reach capture in furnished home layouts, including flooring transitions, narrow passages, and fall/recovery episodes. This is a custom capture program for humanoid navigation and reach in furnished homes. We record synchronized RGB-D, IMU, and proprioception across flooring transitions, narrow passages, and balance-critical reach tasks, including controlled fall and recovery sets. Homes pack the clutter, surface changes, and tight spaces that break whole-body policies trained in open or simulated environments. ## What we collect Locomotion across flooring transitions, narrow passages, and reach while balance-critical, capturing both smooth traversal and the recoveries that occupied, cluttered homes demand. ## Sensors and modalities Synchronized RGB-D, IMU, and proprioception, time-aligned through our [multi-sensor synchronization service](/services/multi-sensor-sync) so whole-body state and footing correspond exactly. ## How capture works A pilot validates rig and consent workflows, then capture scales across home layouts following our [robotics data collection](/services/robotics-data-collection) process, with safety spotting for fall/recovery sets. ## QA and metadata Episodes carry layout type, surface, passage width, reach height, and outcome tags. QA covers sync, calibration, and metadata completeness against agreed acceptance criteria. ## Who it is for [Humanoid robotics](/industries/humanoid-robotics) and home-robotics teams targeting domestic generalization. Complements [humanoid stair ascent and descent](/datasets/humanoid-stair-ascend) for full whole-body locomotion coverage. --- ## Humanoid stair ascent and descent URL: https://www.operantdata.com/datasets/humanoid-stair-ascend Humanoid stair climbing in public and residential stairs. Stair locomotion capture with balance-critical recovery data across varied tread dimensions and handrail availability. This is a custom capture program for humanoid stair ascent and descent across public and residential staircases. We record synchronized RGB-D, IMU, and proprioception through balance-critical footsteps, including stumbles and recoveries captured under safety protocols. Stairs concentrate the contact, timing, and balance failures that simulation struggles to reproduce, which is exactly where humanoid locomotion policies tend to break. ## What we collect Ascent and descent across varying tread dimensions, surface materials, and handrail availability, including clean climbs alongside the stumbles and recoveries that determine fall risk. ## Sensors and modalities Synchronized RGB-D, IMU, and proprioception, time-aligned through our [multi-sensor synchronization service](/services/multi-sensor-sync) so whole-body state corresponds to foot placement and contact. ## How capture works A pilot validates rig placement and episode boundaries, then capture scales across staircases and conditions following our [robotics data collection](/services/robotics-data-collection) workflow, with safety spotting throughout. ## QA and metadata Episodes carry tread geometry, surface, lighting, handrail tags, and outcome (clean, stumble, recovery). QA covers sync, calibration, and label completeness against agreed criteria. ## Who it is for [Humanoid robotics](/industries/humanoid-robotics) teams training locomotion and balance control. The recovery focus relies on deliberate [edge-case data collection](/services/edge-case-scenarios), and the program pairs with [humanoid home navigation](/datasets/humanoid-home-navigation) for whole-body coverage. --- ## Industrial peg-in-hole with force traces URL: https://www.operantdata.com/datasets/industrial-peg-insertion-force Peg-in-hole insertion with force-torque sensing. Contact-rich insertion capture with force-torque and vision-occlusion labels, including alignment errors and jam events. This is a custom capture program for contact-rich peg-in-hole insertion in factory cells. We record synchronized force-torque, RGB-D, and proprioception through alignment errors, jam events, and successful inserts. Insertion is decided by milliseconds-long contact dynamics that simulation approximates poorly, so force-aligned real capture is what makes these policies reliable on the line. ## What we collect Alignment errors, jam events, and successful inserts with synchronized force and video, capturing the full distribution of contact outcomes rather than only clean insertions. ## Sensors and modalities Force-torque, RGB-D, and proprioception, tightly time-aligned through our [multi-sensor synchronization service](/services/multi-sensor-sync). Sync matters here because contact events are milliseconds long. ## How capture works A pilot validates the force-vision rig and acceptance criteria, then capture scales across parts and tolerances following our [robotics data collection](/services/robotics-data-collection) workflow. ## QA and metadata Episodes carry part geometry, tolerance, contact outcome, max force, and cycle time. QA gates apply the insertion success rates and force thresholds agreed in scope. ## Who it is for [Industrial manipulation](/industries/industrial-manipulation) teams working on contact-rich assembly. Because contact dynamics are where simulation breaks down, this program directly supports [sim-to-real data collection](/services/sim-to-real-gap). --- ## Laboratory pipetting dexterity capture URL: https://www.operantdata.com/datasets/lab-pipetting-dexterity Pipetting and micro-manipulation in lab benches. Precision manipulation capture for lab automation prototypes, with sub-millimeter trajectories and contact/alignment failure labels. This is a custom capture program for precision pipetting and micro-manipulation on lab benches. We record sub-millimeter trajectories with synchronized RGB-D, force-torque, and proprioception, labeling contact and alignment failures. Lab automation lives or dies on fine-motion precision, where small alignment errors ruin a draw, and that resolution is something only careful real-world capture provides. ## What we collect Sub-millimeter motion episodes with failure labels for contact and alignment, spanning clean draws and the near-misses that distinguish a reliable protocol from a fragile one. ## Sensors and modalities RGB-D, force-torque, and proprioception, time-aligned through our [multi-sensor synchronization service](/services/multi-sensor-sync) so fine contact corresponds to motion. ## How capture works Demonstrations are gathered through [teleoperation capture](/services/teleoperation-capture). A pilot validates the precision rig and labeling against your labware, then capture scales across protocols. ## QA and metadata Episodes carry plate format, tip type, protocol step, and failure tags. QA enforces precision and success criteria agreed in scope. ## Who it is for Lab automation and precision manipulation teams, adjacent to [industrial manipulation](/industries/industrial-manipulation) work. Shares dexterity-capture methods with [dual-arm kitchen manipulation](/datasets/dual-arm-kitchen-manipulation). --- ## Outdoor rough-terrain locomotion URL: https://www.operantdata.com/datasets/outdoor-rough-terrain-locomotion Rough-terrain legged locomotion outdoors. Legged locomotion capture on uneven outdoor surfaces with synchronized IMU, vision, and LiDAR, including slip and recovery behaviors. This is a custom capture program for legged locomotion on uneven outdoor terrain. We record synchronized RGB-D, IMU, and LiDAR across gravel, grass, and wet surfaces, deliberately capturing slip, trip, and recovery behaviors. Outdoor footing dynamics, loose substrate, slopes, and surface change, are notoriously hard to simulate, and they are where real locomotion policies lose their footing. ## What we collect Slip, trip, and recovery behaviors on gravel, grass, and wet surfaces, capturing both stable gaits and the disturbances that test balance recovery. ## Sensors and modalities Synchronized RGB-D, IMU, and LiDAR, time-aligned through our [multi-sensor synchronization service](/services/multi-sensor-sync) so terrain perception and body state correspond. ## How capture works A pilot validates rig and terrain labeling, then capture scales across surfaces and conditions following our [robotics data collection](/services/robotics-data-collection) workflow. ## QA and metadata Episodes carry surface type, slope, weather, and outcome tags. QA covers sync, calibration, and label completeness against agreed acceptance criteria. ## Who it is for Legged robotics teams, including [humanoid robotics](/industries/humanoid-robotics) groups, that need real terrain data. Directly supports [sim-to-real data collection](/services/sim-to-real-gap) by supplying the outdoor conditions simulation gets wrong. --- ## Retail shelf restocking teleoperation URL: https://www.operantdata.com/datasets/retail-shelf-restock Retail shelf restocking via teleoperation. Shelf-facing pick-and-place capture with SKU diversity in retail aisles, including occlusion and crowded-shelf episodes. This is a custom capture program for teleoperated retail shelf restocking across real store aisles. Operators demonstrate facing and placement across diverse SKUs while we record synchronized RGB-D and gripper state, tagged with occlusion and shelf-layout metadata. It targets the visual clutter and assortment variety of retail, where packaging graphics, crowded shelves, and tight facings break policies trained on clean benches. ## What we collect Facing alignment, occlusion from packaging graphics, and crowded shelf layouts, captured across a diverse SKU set with both clean placements and the misfacings that retail conditions cause. ## Sensors and modalities Multi-camera RGB-D and proprioception, time-aligned through our [multi-sensor synchronization service](/services/multi-sensor-sync), with calibration files delivered per program. ## How capture works Most demonstrations are gathered through [teleoperation capture](/services/teleoperation-capture). A pilot validates the rig and labeling, then capture scales across SKUs, shelves, and lighting. ## QA and metadata Episodes carry SKU class, shelf position, occlusion tags, and placement outcome. QA covers sync, calibration, and metadata completeness against agreed acceptance criteria. ## Who it is for Retail robotics teams whose pick-and-place policies must handle real assortment and clutter, adjacent to [warehouse automation](/industries/warehouse-automation) work. Pairs with the [defective SKU pick failure library](/datasets/defective-sku-pick-failures) for failure-aware evaluation. --- ## Warehouse conveyor handoff episodes URL: https://www.operantdata.com/datasets/warehouse-conveyor-handoff Robot-to-conveyor handoff in distribution centers. Robot-to-conveyor transfer capture in distribution centers with timing-critical grasp and line-speed metadata. This is a custom capture program for robot-to-conveyor handoffs in distribution centers, where timing is everything. We record approach, grasp, place, and miss episodes across line-speed variants with synchronized sensors and handoff-outcome labels. It targets the timing-critical transfers that look trivial in a lab but break under real line speeds, tote variation, and operational pressure. ## What we collect Approach, grasp, place, and miss episodes with line-speed variants, capturing both successful transfers and the mistimed or dropped handoffs that determine real reliability. ## Sensors and modalities Multi-camera RGB-D and proprioception, time-aligned through our [multi-sensor synchronization service](/services/multi-sensor-sync). Tight synchronization matters here because handoff events are milliseconds long. ## How capture works A pilot validates rig placement and timing labels, then capture scales across line speeds and shifts. The program follows our standard [robotics data collection](/services/robotics-data-collection) workflow with handoff-success QA gates. ## QA and metadata Each episode is tagged with line speed, handoff outcome, and optional WMS hooks for order and tote IDs. QA covers sync tolerances, calibration, and metadata completeness against agreed acceptance criteria. ## Who it is for Distribution and fulfillment teams working on [warehouse automation](/industries/warehouse-automation) where throughput depends on reliable transfers. Complements [warehouse pallet pick teleoperation](/datasets/warehouse-pallet-pick-teleoperation) for full pick-to-line coverage. --- ## Warehouse pallet pick teleoperation capture URL: https://www.operantdata.com/datasets/warehouse-pallet-pick-teleoperation Teleoperated pallet picking in operational warehouse aisles. Teleoperated pallet picking with multi-camera RGB-D in live warehouse aisles, captured with grasp-outcome and SKU metadata. This is a custom capture program for teleoperated pallet picking in operational warehouse aisles. Human operators demonstrate picks across varied SKUs and stack heights while we record synchronized multi-camera RGB-D and gripper state, tagged with grasp outcomes and SKU metadata. It exists because simulation rarely reproduces the packaging variance, occlusion, and lighting of a real warehouse, where most pick policies actually fail. ## What we collect Human-guided pallet picks with synchronized depth and gripper state across varied SKUs, stack heights, and aisle positions. Episodes capture clean picks alongside slips, regrasps, and misses so your evaluation reflects real pick success, not a sanitized subset. ## Sensors and modalities Multi-camera RGB-D plus proprioception and gripper state, time-aligned through our [multi-sensor synchronization service](/services/multi-sensor-sync) so observations and actions correspond exactly. Calibration files ship with every program. ## How capture works A pilot week validates the camera rig and labeling schema, then capture scales across shifts with grasp-success QA. The workflow follows our standard [robotics data collection](/services/robotics-data-collection) process: scope, pilot, scale, handoff. ## QA and metadata Each episode carries SKU class, grasp outcome, aisle ID, stack height, and lighting tags. QA gates cover time-sync tolerances, calibration checks, and metadata completeness against the acceptance criteria agreed during scoping. ## Who it is for Teams improving pick policies where simulation misses real packaging variance, particularly groups working on [warehouse automation](/industries/warehouse-automation) and imitation-learning pipelines. Pairs naturally with [warehouse conveyor handoff](/datasets/warehouse-conveyor-handoff) capture for end-to-end material flow. --- # Guides and insights ## Metrics that prove sim-to-real gap closing URL: https://www.operantdata.com/blog/sim-to-real-metrics Track deployment KPIs on targeted real slices before declaring sim-to-real victory. Pair sim training with real eval cells scoped to known mismatch domains, lighting, contact, wear. Operant fills those cells with measured real episodes. --- ## Teleoperation capture best practices for imitation learning URL: https://www.operantdata.com/blog/teleoperation-best-practices Small teleop pilot mistakes become expensive at scale, calibrate early. Lock camera extrinsics, define episode boundaries, and log operator IDs for diversity tracking. See our [teleoperation capture service](/services/teleoperation-capture) for program details. --- ## Building evaluation benchmarks from real-world tails URL: https://www.operantdata.com/blog/eval-benchmarks-real-world Benchmarks copied from training distributions hide failures you will see in production. Construct eval slices for weather, packaging defects, and human interference, with explicit sampling ratios. Custom capture lets you target those slices deliberately. --- ## Privacy and consent for on-site robotics capture URL: https://www.operantdata.com/blog/privacy-consent-on-site On-site capture needs clear consent, minimization, and transfer controls, not afterthoughts. Align with legal on identifiable imagery, retention windows, and on-device masking when required. Scope privacy requirements in the same workshop where you define sensors and volumes. --- ## Running robotics data collection like a production pipeline URL: https://www.operantdata.com/blog/collection-ops-playbook Pilots, QA gates, and operator runbooks turn one-off demos into reliable data supply. Treat capture like CI: versioned configs, daily QA samples, and rollback when sync drift exceeds thresholds. Operant documents runbooks so your team can reproduce conditions months later. --- ## What "good" physical AI data looks like URL: https://www.operantdata.com/blog/physical-ai-data-quality Episode volume alone does not predict deployment success, sync, calibration, and tail coverage matter. Good physical AI datasets are auditable: time synchronization proven, sensor extrinsics stable, and tail behaviors represented at known rates. Teams should define acceptance tests before scale, not after terabytes land in object storage. Book a discovery call from our homepage to align quality bars with your training stack. --- # FAQ ## How is Operant different from buying an open dataset? Open datasets rarely match your robot, sensors, or deployment environment. Operant designs capture around your evaluation goals, not a fixed catalog download. --- ## Do we own the data? Yes. Engagements grant your organization ownership of captured data with terms documented during scoping. --- ## What sensors do you support? RGB-D, LiDAR, IMU, force/torque, and proprioceptive streams are common. We finalize the sensor list on your discovery call. --- ## How long does a collection program take? Pilots often run 2–4 weeks. Production scales over months depending on diversity targets and geographies. --- ## What happens on a discovery call? We review your stack, behaviors, sensors, volume, and quality bar, then outline a pilot and commercial framework. --- ## How do you handle privacy and consent on site? We follow your policies and applicable regulations with consent workflows and secure transfer agreed before capture. --- ## What deliverables and formats do you provide? Time-synchronized logs, calibration files, metadata, and optional labels in formats your training pipeline expects. --- ## Who is Operant for? Robotics and physical AI teams with defined milestones who need auditable real-world data, not generic bulk downloads. --- ## How is pricing structured? Custom projects are scoped fixed-fee or milestone-based after discovery, aligned to environments, sensors, and volume tiers. --- ## What geographies can you operate in? We run programs in North America and Europe by default; other regions are available with lead time for permits and staffing. --- ## Do you provide labeling and annotation? Yes, label schemas, QA, and tooling integrations can be included when defined in scope. --- ## What do we need for a pilot? Target behaviors, sensor list, a staging environment, and an engineering point of contact for format validation. --- ## How do you QA captured data? Automated sync checks, sample review gates, and acceptance tests tied to the statement of work before scale. --- ## Will our project stay confidential? NDAs and restricted-access delivery are standard for commercial robotics programs. --- ## Can data integrate with our sim pipeline? We deliver metadata and formats compatible with sim-to-real workflows and can align episode IDs with your sim domains. --- ## Do you support ongoing collection after handoff? Yes, retainer-style programs maintain diversity and refresh tail behaviors as your product evolves. --- ## Is there a minimum engagement size? Pilots are designed to be meaningful but bounded; production scale follows proven pilot metrics. --- ## Do you provide robots or only operators? We typically capture on client or lessor hardware; operator staffing and rigging are part of scope. --- ## Can you target specific failure modes? Yes, failure taxonomy and sampling ratios are defined explicitly for eval-oriented programs. --- ## Do you support teleoperation programs? Teleoperation and demonstration capture are core offerings with multi-camera synchronization. --- ## Is Operant a dataset marketplace? No. Operant runs custom capture programs for your environments and robots, we are not a broker or off-the-shelf catalog. ---