HABIT Policy Training

Training code for the two vision-language-action (VLA) policies fine-tuned on the HABIT dataset — a large-scale bimanual, human-present robot manipulation dataset in LeRobot v2 format.

This repository packages two independent, self-contained training setups, each vendoring its upstream policy codebase and adding a thin HABIT-specific configuration layer. Both have been adapted to run on a single local machine (no cluster/orchestration dependencies).

Model	Directory	Upstream	HABIT entrypoint
π0.5 (OpenPI)	PI05/	openpi	python -m local_training.train_local samples/envs/.env.habit
GR00T N1.6	Gr00t-N1.6/	Isaac GR00T	bash examples/HABIT/finetune_habit_1gpu.sh

Each directory has its own README with full setup and usage instructions:

• PI05/README.md — OpenPI (π0 / π0.5 / π0-FAST) local training.
• Gr00t-N1.6/README.md — GR00T N1.6 fine-tuning (see the Fine-tune HABIT section).

---

The HABIT data layout both models consume

HABIT is a bimanual Franka Research 3 dataset. Both training setups read the same LeRobot v2 fields:

• State (14D): per-arm end-effector Cartesian position (xyz 3D + rotation 3D)
  • gripper (1D), for the left and right arm.
• Action (14D): per-arm end-effector delta action (7D [xyz 3D, rotation 3D, gripper 1D]), left and right.
• Cameras (robot-side): front_view, left_wrist_view, right_wrist_view.

The full HABIT dataset additionally ships two human-side camera streams (human_front_view, exo_view); the policy-training configs here use only the three robot-side views, matching the experiments in the paper.

The two policies expose these fields differently:

• π0.5 maps LeRobot columns to OpenPI inputs via env vars in PI05/samples/envs/.env.habit (STATE_FIELDS, ACTION_FIELDS, IMAGE_KEYS). It can auto-download a subset of configinc/HABIT from the Hub.
• GR00T reads a local LeRobot v2 directory via --dataset_path and maps fields through Gr00t-N1.6/examples/HABIT/habit_config.py. Download the dataset yourself first (e.g. git clone / huggingface-cli download configinc/HABIT) and point --dataset_path at the resulting directory.

---

Quick start

π0.5 (OpenPI) — auto-downloads a 10-episode HABIT subset

cd PI05
cd openpi && uv sync && cd ..          # install OpenPI deps once
python -m local_training.train_local samples/envs/.env.habit

This stages 10 episodes of the HABIT sample subset under ~/.cache/openpi-local/, pulls the public π0.5 base weights, and runs a 200-step fine-tune. See PI05/README.md for scaling to the full dataset and adapting to other LeRobot v2 datasets.

GR00T N1.6 — point it at a local HABIT checkout

cd Gr00t-N1.6
# install deps per Gr00t-N1.6/README.md (uv / pip), then:
# edit --dataset_path in the script to your local HABIT directory
bash examples/HABIT/finetune_habit_1gpu.sh      # single GPU
bash examples/HABIT/finetune_habit_1node.sh     # 8 GPUs, single node (torchrun)

Checkpoints are written to Gr00t-N1.6/outputs/habit_finetune/ by default.

---

Repository notes

• Vendored upstreams. PI05/openpi/ and Gr00t-N1.6/ contain their respective upstream sources under their original licenses (Apache 2.0). The HABIT-specific additions are limited to PI05/local_training/, PI05/samples/, and Gr00t-N1.6/examples/HABIT/.
• Local-only. Both entrypoints run on one machine (single- or multi-GPU via torchrun); there are no cloud/cluster orchestration dependencies in the training path.
• License. Each vendored codebase retains its upstream license; the HABIT integration code is released under the same terms.