rskill-pi05-franka_panda-libero_spatial-int8

OpenRAL rSkill — π0.5 (3 B PaliGemma backbone, flow-matching action head) finetuned on the LIBERO benchmark, loaded bf16 and LLM.int8-quantized in place so it runs on an 8 GB consumer GPU. Non-commercial weights (Physical Intelligence permissive research license).

⚠ License gate — π0.5 weights are not Apache-2.0. The OpenRAL loader pins commercial_use_allowed: false; commercial deployment requires a separate agreement with Physical Intelligence (see CLAUDE.md §7.4 / Operating Principle 9). The loader requires OPENRAL_ALLOW_NONCOMMERCIAL=1 (or the --non-commercial flag on openral skill install) to activate this skill.

Why int8 (and not NF4)

π0.5's 3.4 B backbone does not fit bf16 on an 8 GB card — bf16 peaks at 8.06 GB of device memory, and an 8 GB GPU has less than that free once the driver takes its cut. NF4 4-bit fits but is too lossy for this backbone: it scores 0/5 on libero_spatial, destroying the policy outright. LLM.int8 (bitsandbytes Linear8bitLt on every Linear ≥ 4 M weight elements) fits at 6.35 GB and keeps the policy working. Quantization runs at load from the bf16 checkpoint (no prequantized pack).

…but int8 is not free — prefer bf16 if it fits

int8 is a memory fit. It costs accuracy and speed, and neither cost was recorded here before. Measured on a GB10 — libero_spatial tasks 0/1/2, seed 42, max_steps=220, n=50 per task per arm (300 episodes), same episodes and protocol, the only difference being quantization.dtype:

dtype	task 0	task 1	task 2	Total	Mean step latency	Peak device memory
bf16	98%	94%	94%	143/150 = 95%	9.6 ms	8.06 GB
int8	76%	22%	82%	93/150 = 62%	24.8 ms	6.35 GB

int8 costs ~33 points of task success (Fisher exact p = 4e-13) and runs ~2.6× slower per step — LLM.int8's mixed-precision decomposition is slower than a plain bf16 GEMM.

Read the per-task columns, not just the total. bf16 is steady (98/94/94); int8 swings between 22% and 82%. The damage is task-dependent, so a single-task spot-check can land on 82% and look acceptable while task 1 is nearly non-functional. That variance is the real hazard — more than the average.

What int8 buys is 1.7 GB, and on an 8 GB card that is decisive: bf16 will not load at all, and 62% beats not running.

On any host with ≳9 GB of VRAM, use bf16. No manifest edit is needed:

openral sim run --config scenes/sim/libero_spatial.yaml \
    --rskill rskills/pi05-libero-int8 --vla-extra dtype=bf16

Faster still on NVIDIA hardware with TensorRT: an FP8 engine reaches 78.9 ms per policy call against bf16 eager's 190 ms and scores 49/50 — identical to bf16, so FP8 costs no task success at all. See the openral-pro TensorRT runtime (OPENRAL_PI05_TRT=1, OPENRAL_PI05_TRT_PRECISION=fp8_bf16_attn).

Scope: one task of LIBERO-Spatial's ten, so treat 98%/76% as a like-for-like comparison rather than a suite score. The older "0.5–0.7 across 10-episode runs" figure is not contradicted so much as too noisy to use — at n=10 the binomial SD is ~0.16, wide enough to cover both arms above.

Quick start

import os
os.environ["OPENRAL_ALLOW_NONCOMMERCIAL"] = "1"

from openral_rskill.loader import rSkill
pkg = rSkill.from_yaml("rskills/pi05-libero-int8/rskill.yaml")

# CLI (will prompt to accept the non-permissive license unless --yes is passed):
uv run openral skill install OpenRAL/rskill-pi05-franka_panda-libero_spatial-int8 --non-commercial --yes

# LIBERO closed-loop sim (int8 fits 8 GB):
PYTORCH_ALLOC_CONF=expandable_segments:True \
  openral benchmark scene --config scenes/benchmark/libero_spatial.yaml \
    --rskill rskills/pi05-libero-int8

Upstream model

Field	Value
Source repo	lerobot/pi05_libero_finetuned_v044
Base model	lerobot/pi05_base
Paper	arxiv:2410.24164 — π0: A Vision-Language-Action Flow Model for General Robot Control
Architecture	PaliGemma 3 B backbone + flow-matching action head
Code license	Apache-2.0
Weights license	Physical Intelligence permissive research (non-commercial)
Parameters	~3 B
Benchmark	LIBERO

Supported robots

Robot	Embodiment tag	Status	Notes
Franka Panda (LIBERO sim)	franka_panda	✓ matches	Native training embodiment.
Other 7-DoF arms	—	requires obs-format adapter	State dim is 8-D LIBERO-style.

Sensors required

Key	Modality	Min resolution
observation.images.camera1	RGB	224 × 224
observation.images.camera2	RGB	224 × 224

Manifest summary

Field	Value
name	OpenRAL/rskill-pi05-franka_panda-libero_spatial-int8
version	0.1.0
license	permissive_research
role	s1
runtime / quantization.dtype	pytorch / int8
weights_uri	hf://lerobot/pi05_libero_finetuned_v044
latency_budget.per_chunk_ms	200 ms (3 B model is heavier than SmolVLA)
commercial_use_allowed	false

Full schema: openral_core.RSkillManifest.

Evaluation

eval/scene_libero_spatial.json holds the locally-reproduced result: libero_spatial = 0.50 (5/10 episodes), int8, RTX 4070 8 GB (a separate 10-episode run scored 0.70; the policy sits around 0.5–0.7).

⚠️ n=10 is too small to compare anything with. Its binomial SD is ~0.16, so that 0.50 and the 0.76 measured at n=50 above are the same result, not a regression or an improvement. Prefer the n=50 numbers in Why int8 when deciding a dtype, and treat the recorded 0.50 as provenance for the committed eval artifact rather than as this policy's success rate.

Re-run with:

PYTORCH_ALLOC_CONF=expandable_segments:True OPENRAL_ALLOW_NONCOMMERCIAL=1 \
  openral benchmark scene --config scenes/benchmark/libero_spatial.yaml \
    --rskill rskills/pi05-libero-int8 --n-episodes 10 --write-eval

License

This rSkill package (rskill.yaml, README.md, eval/scene_libero_spatial.json) is Apache-2.0. The wrapped weights are released under Physical Intelligence's permissive research license — review the upstream license file before any deployment beyond research.

See also

• rskills/smolvla-libero/README.md — Apache-2.0 LIBERO alternative.
• docs/reference/vla_compatibility.md §3.1 — VLA × Robot × Sim matrix.
• CLAUDE.md §7.4 — VLA license matrix and install-time guard rules.