Gurukul
Physical AI on a real robot arm — working the full loop on hardware, not in simulation: data collection, training, evaluation, serving, and observability.
Gurukul is a set of hands-on physical-AI projects on a Seeed Studio SO-ARM101 Pro (a leader and a follower arm) using Hugging Face LeRobot. The aim is to run the entire loop on real hardware rather than in a simulator — collecting data, training a policy, evaluating it on the physical arm, and then serving it with observability.
The naming is the setup: the leader arm you drive by hand is guru, the follower that learns to imitate is shishya — teacher and student.
The run
Result
Project 01 is complete. An ACT policy trained on 45 teleoperated demonstrations picks up a stuffed toy and drops it into a bowl in 10 of 20 trials on the real arm — toy from three marked spots, bowl at a fixed position, wrist and front cameras, a 45-second limit, evaluated under lighting that differed from the recording sessions. It is a baseline, honestly scored on hardware, not a cherry-picked best run.
The loop
- Collect 45 teleoperated demonstrations on the real arm — the leader arm (guru) drives the follower (shishya), recorded from a wrist camera and a fixed front camera at 640×480, 30 fps.
- Train An ACT policy trained from the recorded dataset on Hugging Face Jobs (a10g-large); the dataset and checkpoints live on the HF Hub.
- Evaluate 20 trials on the physical arm: toy from three marked spots into a fixed bowl, a 45-second limit, scored in room lighting that deliberately differed from the recording sessions.
- Serve & observe Projects 02–03 (next): policy serving with observability, and Ray-based data curation.