rskill-pi05-so100

OpenRAL rSkill — Physical Intelligence π0.5 (3.4 B-param PaliGemma-backbone VLA), packaged for the SO-100 follower arm.

This package wraps lerobot/pi05_base with a rskill.yaml manifest. It does not copy model weights.

⚠ License (CLAUDE.md §7.4)

Component	License
Code (rskill.yaml, README.md)	Apache-2.0
Weights (lerobot/pi05_base)	Physical Intelligence "permissive research" — not full Apache-2.0

Commercial deployment is prohibited without a separate agreement with Physical Intelligence. The OpenRAL loader enforces this at load time:

# rskill.yaml
license: "permissive_research"
commercial_use_allowed: false

To load this skill, you must explicitly acknowledge non-commercial use via the OPENRAL_ALLOW_NONCOMMERCIAL=1 environment variable, or pass commercial_use=False to rSkill.from_pretrained(...).

Upstream model

Field	Value
Source repo	lerobot/pi05_base
Paper	arxiv:2410.24164 — π0.5: A Vision-Language-Action Model with Open-World Generalization
Backbone	PaliGemma (Gemma 2B) — ~3.4 B params total
Action chunk	50
Benchmark	none — base checkpoint, no task-specific finetune

Memory note. FP32 weights ≈ 13.6 GiB (OOM on an 8 GiB GPU). This manifest pins quantization.dtype = bf16 so weights load at ≈ 6.8 GiB.

Supported robots

Robot	Embodiment tag	Status	Notes
SO-100 follower arm (real or SO100DigitalTwin)	so100_follower, lerobot	✓ IO contract	6-DoF, single RGB stream

Sensors required

Key	Type	Resolution	Format
observation.images.camera1	RGB camera	224 × 224 (min)	float32

The exact camera layout depends on the loaded checkpoint; the rSkill declares one canonical RGB stream as the contractual minimum. The eval adapter pulls additional camera keys from the loaded config at runtime.

Manifest summary

Field	Value
name	OpenRAL/rskill-pi05-so100
version	0.1.0
license	permissive_research
role	s1
embodiment_tags	so100_follower, lerobot
runtime / quantization.dtype	pytorch / bf16
weights_uri	hf://lerobot/pi05_base
latency_budget.per_chunk_ms	2 500 ms (PaliGemma forward dominates; bf16 reference-host pending)
latency_budget.warmup_ms	30 000 ms
latency_budget.load_ms	60 000 ms
commercial_use_allowed	false

Full schema: openral_core.RSkillManifest — python/core/src/openral_core/schemas.py.

Reproduction

git clone https://github.com/OpenRAL/openral && cd OpenRAL
just bootstrap && uv sync --all-packages --group sim

# Manifest validation only (no GPU, no weights):
uv run pytest tests/unit/test_rskill_loader.py -v

# Closed-loop sim against LIBERO Franka via `openral sim run` (≥8 GB VRAM required;
# requires non-commercial acknowledgement):
OPENRAL_ALLOW_NONCOMMERCIAL=1 just sim-pi05-libero
# which runs:
#     openral sim run --config scenes/benchmarks/pi05_libero_spatial.yaml --save-video

See also

• robots/so100_follower/README.md — RobotDescription manifest.
• rskills/pi05-libero-nf4/README.md — π0.5 LIBERO finetune.
• docs/reference/vla_compatibility.md — VLA × Robot × Sim matrix.