Teleoperation Data Collection
SVRC operates a multi-station teleoperation facility with 50+ trained operators and three collection modalities. We collect the high-quality demonstration data your robot policies need — from 20-episode pilots to 10,000+ episode production campaigns.
Collection Modalities
VR Teleoperation
Meta Quest 3 and Apple Vision Pro headsets mapped to robot end-effectors. 6-DOF wrist control with grip trigger. Ideal for single-arm manipulation, pick-and-place, and tasks requiring spatial awareness. Operators see the workspace through robot-mounted cameras with low-latency passthrough.
Leader-Follower (Bilateral)
ALOHA-style paired arms where a kinematically identical leader arm drives the follower in real time. Records 50 Hz joint-level data with sub-millimeter correspondence. The gold standard for bimanual manipulation — threading, assembly, packing, and contact-rich tasks that require precise force control.
Glove-Based Control
SenseGlove Nova and BrainCo haptic gloves for dexterous hand teleoperation. Per-finger joint tracking at 90 Hz with haptic feedback. Used for tasks requiring individual finger control — tool use, deformable object manipulation, and human-like grasp strategies on multi-fingered end-effectors.
Supported Robots
We maintain collection stations for OpenArm, Franka FR3, UR5e/UR10e, xArm 6/7, Kinova Gen3, WidowX/ViperX (ALOHA), and Mobile ALOHA. If your robot runs ROS 2, we can integrate it into our collection infrastructure within 3–5 business days. Browse our hardware catalog for purchase or lease options.
Delivery Formats
Every dataset is delivered in your target format, validated against your training pipeline before handoff:
- HDF5 — ACT / ALOHA compatible, per-episode files with synchronized camera, state, and action streams
- RLDS (TFRecord) — Open X-Embodiment and Octo compatible, ready for cross-embodiment pretraining
- LeRobot Parquet — Hugging Face Hub ready, streaming-compatible, with standardized column schema
- Custom — RoboMimic HDF5, RLBench, or your own schema
Quality Assurance
Every episode passes automated QA before delivery: trajectory smoothness scoring, grasp success verification, camera timestamp synchronization checks (<5 ms tolerance), and format schema validation. Failed episodes are re-collected, not patched. Operators are proficiency-tested on each task before collection begins.