Φ — Physical Hardware Intelligence

An open, reproducible robot-learning pipeline — hardware to trained policy, documented end to end.

_{Our SO-101, front camera, from phi_so101_cubes_cylinder_v1 — real recorded data, not a render.}

---

What this is

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.

Quickstart

git clone https://github.com/physical-hardware-intelligence/phi.git && cd phi
make install                 # env + package + verify   (ENV=cuda on a GPU box)
conda activate phi
make doctor                  # tells you what is still wrong

New here? Read docs/00-setup-your-laptop.md instead — it covers prerequisites, what each package is for, how to verify the install actually worked, and the failure modes that are silent if you skip the check.

For the shape of the whole project: docs/00-overview.md.

Where do I go?

I want to…	Go here
Set up your laptop	Setup your laptop
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 a trained policy on the arm	Deployment — inference
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

The member ladder

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

What's in here

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

Status

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/.

License & credits

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.