Rudra Robotics
Rudra Robotics provides a unified data capture-to-evaluation pipeline for robot foundation models, recording human work through egocentric headsets, handheld grippers, and teleoperation—all synchronized to a single clock with sub-millisecond precision. The company normalizes and annotates data into model-ready formats like LeRobot v3 and RLDS, then benchmarks trained policies and closes performance gaps with targeted data collection and reinforcement learning fine-tuning.
- Robotics
- Artificial Intelligence
- Hardware
Funding
Founders
Founder details are not available yet.
Product
Problem
Robot manipulation models require large volumes of high-fidelity, task-specific data to learn real-world skills, but existing collection methods are fragmented across incompatible rigs, formats, and timebases. This makes it difficult for robotics teams to obtain datasets that accurately capture the sensory and action streams needed to train effective policies, forcing them to stitch together inconsistent recordings or run costly in-house collection efforts.
Solution
Field Log provides a unified data capture-to-evaluation pipeline for robot foundation models. The company records people performing real work using three synchronized modalities: egocentric headsets with six fisheye cameras and hand tracking, handheld UMI-style grippers with jaw encoders, and teleoperated robot demonstrations. All streams are aligned to a single clock with sub-millisecond precision, then normalized, annotated, and delivered in model-ready formats like LeRobot v3 and RLDS. After a model is trained, Field Log benchmarks its performance, builds a failure taxonomy, and closes identified gaps with targeted data collection and reinforcement learning fine-tuning.
Target Audience
Primary customers are robotics teams and AI labs training manipulation policies or robot foundation models, including those building general-purpose humanoid or arm-based systems for factory, kitchen, and home environments.
Features
- Egocentric headset with six synchronized fisheye cameras, 21-joint hand skeleton tracking via VIO, and a 500 Hz IMU to capture the human view of work
- Bimanual handheld UMI-style grippers with jaw encoders and per-unit pose tracking, placing recordings directly in the action space a manipulation policy learns
- Teleoperation capture using trained operators driving real robot arms and humanoids to produce robot-native trajectories
- Unified timebase synchronization across all capture modalities with ≤1 ms precision
- Data normalization pipeline that calibrates, reconstructs poses, segments episodes, blurs faces, and annotates hand pose, object/contact state, task language, and success/failure labels
- Delivery in LeRobot v3 or RLDS formats with synchronized camera streams, poses, IMU, calibration, and per-frame annotations
- Post-training evaluation loop with task suites, failure taxonomy generation, and targeted data collection plus RL fine-tuning
- Partner network reaching more than 80,000 factories and a workforce of 1.2 million people across India, Indonesia, and China for large-scale collection