An open, reproducible robot-learning pipeline — hardware to trained policy, documented end to end.
License Built on LeRobot Robot Docs
SO-101 pick and place, front camera
Our SO-101, front camera, from phi_so101_cubes_cylinder_v1 — real recorded data, not a render.
The robotics SIG at Northeastern University, Silicon Valley. Everything we build — robots, datasets, policies, training runs, evaluations — lives in this one repo, versioned and documented so a new member can get to a trained policy without starting from scratch.
We don't reinvent the engine. LeRobot does the driving; Φ adds the curriculum, curation, evaluation protocol, and reproducibility layer on top.
git clone <this-repo-url> phi && cd phi conda env create -f env/environment.mac.yml # or environment.cuda.yml on a GPU box conda activate phi pip install -e . make help
Then open docs/00-overview.md .
| I want to... | Go here |
|---|---|
| Understand the whole thing | Overview |
| Build or buy the arm | Hardware & build |
| Get an arm running | Setup & bring-up |
| Record a dataset | Teleop & data |
| Train a policy | Training — the policy zoo |
| Score a policy honestly | Evaluation protocol |
| Run it on the robot / edge | Deployment |
| Train on the cluster | Explorer HPC |
| Do kinematics in sim | Simulation (MuJoCo) |
| Know why a policy works | Theory notes |
| Something is broken | Troubleshooting |
| Contribute | CONTRIBUTING |
| Level | You can... | Start here |
|---|---|---|
| L0 Onboard | replay a recorded episode | overview |
| L1 Operator | calibrate, teleoperate, record a dataset | 02-setup |
| L2 Trainer | train a policy on your data and score it | training · evaluation |
| L3 Contributor | add a task or config, close a good-first-issue | tasks/TEMPLATE |
| L4 Researcher | run a new experiment and write it up | experiments/TEMPLATE |
| L5 Maintainer | own a module, review PRs | CONTRIBUTING |
docs/ the curriculum (mkdocs site)
simulation/ MuJoCo + SO-101 kinematics — runs on a Mac
src/phi/ thin tooling over LeRobot
configs/ pinned, seeded run configs
datasets/ dataset cards — data lives on the HF Hub
models/ model cards, one per checkpoint
experiments/ dated write-ups of every run
tasks/ task specs + eval rubrics
env/ tests/ environments · unit + smoke tests
Phase 1 — training and evaluating on our own arm.
| Datasets recorded | 3 public on the HF Hub, 3-camera (wrist · front · top) |
| Policies trained | ACT, Diffusion Policy (CNN + Transformer), patch-encoder variants |
| Model cards | 6 |
| Experiment write-ups | 9 |
| Rollouts scored on the real arm | 62, against a written rubric |
| Simulation | MuJoCo FK/IK on the SO-101, verified against the model |
Cockpit is a Mac (record + teleop); training runs on a CUDA box or the Explorer cluster.
We publish negative results. Several experiments here record things that did not work, and one carries a correction notice over its original conclusion. That is deliberate — see
experiments/.
Apache-2.0 — see LICENSE. Built on LeRobot and the SO-ARM100/101 hardware project.
Φ is a Student Interest Group at Northeastern University, Silicon Valley. Not affiliated with or branded by the university.