Teleoperation Training Data

Claru operates a distributed network of teleoperation collection stations equipped with ALOHA-style leader-follower systems, 3D SpaceMouse interfaces, and multi-view camera arrays. We support Franka Panda, UR5e, ViperX 300, and custom platforms with standardized calibration across all stations. Our trained teleoperators achieve 30+ successful demonstrations per hour on tabletop tasks, with quality maintained through 45-minute session limits, per-operator qualification protocols, and automated per-episode quality scoring. We deliver in all major formats (RLDS, HDF5, Zarr) with complete metadata including camera calibration, robot URDF, action space definitions, and task specifications. For foundation model training campaigns, we sustain 1,000+ quality demonstrations per week with continuous quality monitoring and format compliance verification.

Teleoperation — where a human remotely controls a robot to perform tasks — is the primary method for collecting high-quality robot demonstration data. Unlike kinesthetic teaching (physically guiding the robot) or autonomous exploration, teleoperation combines human task intelligence with the robot's actual kinematic and dynamic constraints, producing demonstrations that are both naturally varied and physically executable. Every major robot learning dataset of the past three years — RT-1, DROID, Open X-Embodiment, ALOHA — relies on teleoperated demonstrations as the foundational data source.

The choice of teleoperation interface fundamentally shapes the data quality and throughput. Leader-follower systems like ALOHA (Zhao et al., 2023) and GELLO (Wu et al., 2024) offer sub-millimeter positional fidelity and natural bimanual control, but require custom hardware costing $5K-20K per station. VR-based teleoperation using Meta Quest or HTC Vive provides lower entry cost but introduces 20-50 ms latency and kinematic mismatch between human and robot workspaces. Exoskeleton interfaces offer the best kinematic correspondence but cost $50K-200K per station. Each interface type produces demonstrations with subtly different quality characteristics: leader-follower data preserves fine-grained force modulation, VR data captures spatial reasoning well but loses force sensitivity, and exoskeleton data provides the most natural motion profiles but is least scalable.

At scale, the economics of teleoperation data collection are driven by operator throughput and quality. Experienced teleoperators on tabletop tasks typically achieve 20-40 successful demonstrations per hour with leader-follower systems, compared to 10-20 with VR and 5-10 with kinesthetic teaching. Google's RT-1 data collection campaign (Brohan et al., 2022) employed 13 robots with scripted teleoperators over 17 months, collecting 130,000 demonstrations — a benchmark for scale that few organizations can match independently. The DROID dataset (Khazatsky et al., 2024) distributed collection across 13 institutions with standardized Franka Panda arms and 3D SpaceMouse interfaces, producing 76,000 demonstrations in 12 months with consistent quality.

Data quality in teleoperation is not just about sensor fidelity — it is about demonstration naturalness. Policies trained on jerky, hesitant teleoperation perform measurably worse than those trained on fluid demonstrations from experienced operators. Mandlekar et al. (2021) showed that filtering the top 50% of demonstrations by smoothness and efficiency metrics improved Behavioral Cloning success rates by 12-18% across tasks. This makes operator selection, training, and fatigue management as important as hardware selection for dataset quality.

The dominant paradigm in 2024-2026 is to train large visuomotor policies on teleoperated demonstration data. RT-2 (Brohan et al., 2023) fine-tuned a vision-language model on 130K teleoperated episodes and achieved emergent generalization — following novel language instructions for object manipulation without task-specific demonstrations. OpenVLA (Kim et al., 2024) replicated this approach with open-source models and 970K episodes from the Open X-Embodiment dataset, achieving 16.5% absolute improvement over RT-2-X on real robot evaluation. The key finding across all these systems is that data diversity — more tasks, more objects, more environments — matters more than data volume for a single task.

Diffusion Policy (Chi et al., 2023) demonstrated that policy architecture innovation can dramatically reduce the data requirement. By modeling the multimodal action distribution with a denoising diffusion process, Diffusion Policy achieves 85-95% success rates on contact-rich tasks from just 50-200 teleoperated demonstrations — compared to 200-500 demonstrations needed for standard Behavioral Cloning. ACT (Zhao et al., 2023) achieves similar data efficiency through action chunking, predicting sequences of 50-100 future actions at once. Both architectures are now standard choices for consuming teleoperation data.

The DROID dataset (Khazatsky et al., 2024) represents the current best practice for distributed teleoperation collection. Using a standardized hardware setup (Franka Panda + 3 cameras + 3D SpaceMouse) across 13 institutions, DROID collected 76,000 demonstrations covering 564 tasks in 86 environments. The standardization enabled cross-institution data aggregation while the distribution enabled environment diversity no single lab could achieve. Policies trained on DROID show 20% higher success rates on held-out tasks compared to single-institution datasets of equivalent size, validating the diversity hypothesis.

The emerging frontier is scaling teleoperation data to millions of demonstrations for training robot foundation models. Physical Intelligence's pi0 model, Nvidia's GR00T, and Google DeepMind's Gemini Robotics all target million-demonstration-scale training. At this scale, the bottleneck shifts from data quality to data logistics: managing thousands of collection hours across dozens of stations while maintaining format consistency, sensor calibration, and quality standards. This is fundamentally a data operations problem, not a research problem.

Parallelization is the key to economical large-scale teleoperation data. Running multiple collection stations simultaneously with rotating operators can achieve 200-500 demonstrations per day per site. Claru operates teleoperation collection stations across multiple facilities with standardized hardware configurations, enabling parallel collection at throughputs of 1,000+ demonstrations per week while maintaining quality standards. Each station is calibrated to a common specification, ensuring that data from different stations is interchangeable for training.

Data quality in teleoperation collection is maintained through three layers: (1) real-time automated checks for sensor synchronization, episode completeness, and task success; (2) per-session human review of sampled episodes for demonstration quality (smoothness, efficiency, naturalness); (3) dataset-level statistical validation after collection batches complete, including action distribution analysis and per-task success rate verification. Episodes failing quality gates are flagged for re-collection while operators and hardware are still configured, avoiding expensive re-mobilization.

Operator management is the most underappreciated factor in teleoperation data quality. A fresh operator produces measurably smoother demonstrations than one who has been teleoperating for two hours straight. Cognitive fatigue manifests as increased jerkiness, longer episode duration, and higher failure rates. Claru enforces 45-minute collection sessions with 15-minute breaks, rotates operators across tasks to prevent adaptation bias, and tracks per-operator quality metrics to identify when retraining is needed. Operators complete a qualification protocol for each task before production collection begins.

Format standardization is critical for downstream consumption. Claru delivers teleoperation data in RLDS format (compatible with OpenVLA, Octo, RT-X), HDF5 (Diffusion Policy, ACT), or custom formats with full metadata including camera intrinsics/extrinsics, robot URDF, action space definitions, and task specification documents. All data ships with a validation script that confirms format compliance, sensor alignment, and episode completeness before the client begins training.

Claru operates a distributed network of teleoperation collection stations equipped with ALOHA-style leader-follower systems, 3D SpaceMouse interfaces, and multi-view camera arrays. Each station is built to a standardized specification with calibrated multi-view cameras (2-4 per station), hardware-synchronized sensor streams, and validated action space mappings. We support Franka Panda, UR5e, ViperX 300, and custom robot platforms, with data formatted for the target architecture from collection start.

Our operator pipeline ensures consistently high-quality demonstrations. Each operator completes a task-specific qualification protocol before production collection, demonstrating required success rates, trajectory smoothness above threshold, and episode duration within target ranges. During collection, we enforce 45-minute session limits with mandatory breaks, rotate operators across tasks, and run automated quality scoring on every episode. Per-operator dashboards track quality metrics over time, enabling early detection of fatigue or skill regression.

Claru delivers teleoperation data in all major formats — RLDS (OpenVLA, Octo, RT-X), HDF5 (Diffusion Policy, ACT), Zarr, and custom schemas — with complete metadata packages including camera intrinsics and extrinsics, robot URDF files, action space definitions, task specification documents, and per-episode quality scores. Every dataset ships with a validation script and format compliance certificate. For clients building robot foundation models, we support continuous collection campaigns producing 1,000+ quality demonstrations per week at sustained throughput.