xPhi Robotics

Physical-world data & infrastructure for AI. Capture → pipeline → training → evals for robotics & model teams. Human × Humanoid.

[xPhi Robotics @xphirobotics]

Introducing XP-6: a six-camera, head-mounted egocentric capture platform for robot learning.

- 3 stereo pairs, 30 fps per eye
- Up to 180° FOV
- Global shutter + 6-axis IMU
- ±3 ms median cross-module sync

Human-like vision. Real-world data.

#PhysicalAI #Robotics

[xPhi Robotics @xphirobotics]

@LeRobotHF @NVIDIARobotics @bfl_ai LIBERO scores are near the ceiling now (97.9 to 99.0). Is anyone benchmarking these world-action models on real contact-rich tasks where the future-prediction step has to model forces, not just visuals?

[xPhi Robotics @xphirobotics]

@codecopenflow @nvidia A 0.92 correlation between digital-twin and real evals is what makes sim useful for iteration. Do you know if the 6.7% to 63.3% jump held for contact-rich tasks, or mostly for free-space motion?

[xPhi Robotics @xphirobotics]

@SharpaRobotics @nvidia Scissors and glass slides are a tough test for hands. For tasks like that, which signal limits learning more: tactile feedback, or the quality of the human demonstrations the policy starts from?

[xPhi Robotics @xphirobotics]

@Kaweees1 Demos-per-new-task is the right metric. Did you check how demo quality changes the curve, for example clean teleop versus noisier demos? I would expect 50 good demos to beat 200 sloppy ones.

[xPhi Robotics @xphirobotics]

@yuxuank02 Good luck with the posters. For dexterous learning from human data, which bridge has been harder in your experience: retargeting human hands to the robot hand, or getting reliable hand pose from the human video?

[xPhi Robotics @xphirobotics]

@CyberRobooo Hours alone hide the real gap: a lot of raw head-cam footage is unusable until it is cleaned, split into task steps and given hand pose. Do the "effective" hours in those estimates count after QC, or raw capture?

[xPhi Robotics @xphirobotics]

Robot data sources, ranked. Our take:

S: real robot, real task
A: teleop demos, egocentric human video with 3D hands
B: handheld grippers, sim
C: raw web video
D: "more data" with no QC

What would you move? #PhysicalAI #Robotics

[xPhi Robotics @xphirobotics]

@BostonDynamics Using wheels as a controlled sim-to-real probe is a smart way to separate controller fidelity from legged contact noise. Curious which gap was larger: actuator dynamics, or contact with the ground?

[xPhi Robotics @xphirobotics]

@robo_nine A 20% cheaper leader/follower pair means many more people can collect demos at home. Native LeRobot support will matter most for the data format: if everyone records in the same schema, community datasets can be pooled and reused instead of living as one-offs.

[xPhi Robotics @xphirobotics]

@paigeinsf @se3labs @ycombinator Congrats on the launch. In real-world RL the bottleneck is usually reset cost plus how fast you can tell why a try failed. Are you saving head-cam and wrist video per attempt, so failed tries can be triaged and reused as training data?

[xPhi Robotics @xphirobotics]

See it on real clips: 15 head-mounted samples with hand mesh, tracking, camera path and 3D world. https://t.co/yTmZH5Xzst

[xPhi Robotics @xphirobotics]

One second of first-person robot data, unpacked:

120 images (2 sensors x 60 fps, 1920x1200)
600 IMU samples
under 1 microsecond stereo sync

It only becomes training data once hand mesh, camera path, 3D world and task labels are added. What is the first thing your policy pipeline needs from a head-cam clip? #PhysicalAI

[xPhi Robotics @xphirobotics]

@LeRobotHF @lancedb Streaming matters even more for first-person data: one capture session is several synced streams (stereo pair, IMU, depth or hand labels), so downloading whole datasets stops scaling fast. Does 0.7 keep multi-camera windows aligned per worker, or are views sharded separately?

[xPhi Robotics @xphirobotics]

@ZhongyuLi4 Useful, thanks. That crop-to-match trick is why we record the full 210° FOV with IMU embedded in every frame, so you can crop or reproject to the deployment cam afterwards. Have you compared a plain crop against reprojecting to a virtual pinhole for the policy input?

[xPhi Robotics @xphirobotics]

CoRL week question for robot learning teams: what is the biggest bottleneck in your data pipeline right now? Collecting enough demos, cleaning and labeling, hand and contact quality, or the sim-to-real gap? Tell us yours. #PhysicalAI #Robotics

[xPhi Robotics @xphirobotics]

@notmahi Learning dexterity from plain human footage, with no teleop in the loop, is where first-person data gets interesting. Which parts did the human video supply: object motion and intent, or the hand contact itself?

[xPhi Robotics @xphirobotics]

@NVIDIARobotics @RLWRLD_ai @UCBerkeley Cable connecting and small-part assembly are exactly where sim-to-real breaks on contact. Are the new benchmarks scored against real-robot runs too, or only in Isaac Lab-Arena?

[xPhi Robotics @xphirobotics]

@robbyant_brain Cross-sensor transfer is the hard part. Curious how much of the 5,000 hours was paired with vision and human-hand motion, since contact-rich skills usually need that alignment to be useful for policies.