OpenRAL/rskill-pi05-robocasa365-human300-nf4

pi05-robocasa365-human300-nf4

OpenRAL rSkill — pre-quantized nf4 packaging of Physical Intelligence's π₀.₅ (3.4 B PaliGemma backbone) fine-tuned on the RoboCasa365 Human-300 task suite (300 atomic + composite kitchen tasks, 100 demos each) against the PandaMobile embodiment.

Upstream model

Field	Value
Source	robocasa/robocasa365_checkpoints (multitask_learning/75000), converted via tools/openpi_to_lerobot_pi05.py and quantized via tools/quantize_rskill.py.
HF mirror	OpenRAL/rskill-pi05-robocasa365-human300-nf4
Training data	RoboCasa365 Human-300: 300 atomic + composite kitchen tasks, 100 demos each.
Architecture	π₀.₅: 3.4 B PaliGemma backbone + flow-matching action head.
License	Apache-2.0 (code + weights)

Supported robots

Robot	Embodiment tag	Status	Notes
PandaMobile (Franka Panda on a mobile base)	franka_panda	✓ sim	RoboCasa Kitchen default robot; state layout = human300_16d.

Sensors required

Key	Modality	Resolution	Notes
observation.images.robot0_agentview_left_image	RGB	256 × 256	Aliased to the policy's camera1 key via image_preprocessing.aliases.
observation.images.robot0_agentview_right_image	RGB	256 × 256	Aliased to camera2.
observation.images.robot0_eye_in_hand_image	RGB	256 × 256	Aliased to camera3.
observation.state	proprioception	(16,)	human300_16d layout: eef_pos(3) · eef_quat(4) · base_pos(3) · base_rot(4) · gripper(2).

Manifest summary

Field	Value
name	OpenRAL/rskill-pi05-robocasa365-human300-nf4
version	0.1.0
license	apache-2.0
role	s1
embodiment_tags	franka_panda
runtime / quantization.dtype	pytorch / int4 (nf4 / bitsandbytes / bf16 compute)
weights_uri	hf://OpenRAL/rskill-pi05-robocasa365-human300-nf4
chunk_size	50
state_contract	human300_16d named layout
commercial_use_allowed	true

Full schema: openral_core.schemas.RSkillManifest.

Why nf4?

The HF-hosted mirror ships the already-packed nf4 state dict plus a quantization_metadata.json sentinel. The pi05 adapter detects the sentinel, meta-initialises the policy graph (~14 s instead of the ~137 s from_pretrained walk), overlays the prequant state via install_prequantized_linears, and skips the bf16 → nf4 conversion entirely. Warm-up drops to ~20 s on a 4070-mobile (8 GiB).

Running it

OPENRAL_ALLOW_ROBOCASA_ASSETS=1 \
  uv run openral sim run \
    --config scenes/benchmarks/pi05_robocasa_pnp_nf4.yaml \
    --rskill rskill://rskills/pi05-robocasa365-human300-nf4 \
    --view --max-steps 200

The robocasa group conflicts with libero in a single venv; see docs/tutorials/sim/create-a-sim-environment.md ("Level 6: a custom MuJoCo environment via RoboCasa") for the one-time setup (clone robocasa, install robosuite@master, fetch the CC-BY-4.0 kitchen assets).

Image preprocessing

flip_vertical: true — the canonical openpi-robocasa eval pulls images through RoboCasaGymEnv.process_img which applies img[::-1, :, :]. The adapter applies the same flip before forward; the alias map routes the three robosuite cameras into the policy's camera1/2/3 keys.

License

Apache-2.0 — both the wrapping rSkill package (rskill.yaml, README.md) and the wrapped upstream checkpoint (robocasa/robocasa365_checkpoints). Commercial use is allowed (commercial_use_allowed: true).