rskill-xvla-libero

OpenRAL rSkill — xVLA (lerobot ecosystem VLA) finetuned on the LIBERO benchmark. Apache-2.0 weights; locally unverified (TBD checkpoint inspection — see eval/libero.json).

Quick start

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

just sim-xvla-libero --no-run   # validate wiring only
just sim-xvla-libero            # full run (LIBERO sim; needs GPU + MUJOCO_GL)

Upstream model

Field	Value
Source repo	`lerobot/xvla-libero`
Base model	`lerobot/xvla-base`
Paper	TBD
Code license	Apache-2.0
Weights license	Apache-2.0
Benchmark	LIBERO

State dimensions, camera names, normalisation statistics, and the underlying paper are TBD pending checkpoint inspection — see eval/libero.json::status: pending and docs/reference/vla_compatibility.md §3.1.

Supported robots

Robot	Embodiment tag	Status	Notes
Franka Panda (LIBERO sim)	`libero`, `franka_panda`	✓ matches manifest	Native training embodiment (assumed).
Other 7-DoF arms	—	needs adapter	A third 224×224 zero-tensor slot is filled in-process by the runner; it does not correspond to a physical sensor.

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-xvla-libero`
`version`	`0.1.0`
`license`	`apache-2.0`
`role`	`s1`
`runtime` / `quantization.dtype`	`pytorch` / `bf16`
`weights_uri`	`hf://lerobot/xvla-libero`
`latency_budget.per_chunk_ms`	200 ms (conservative — update after local profiling)
`latency_budget.warmup_ms` / `load_ms`	15 000 ms / 60 000 ms
`commercial_use_allowed`	`true`

Full schema: openral_core.RSkillManifest.

Evaluation

eval/libero.json::status is pending — no locally-reproduced numbers and no paper to cite. Do not populate this README's numbers without either a locally-verified run or a precise paper citation with table reference (per CLAUDE.md operating principle 2: truth over plausibility).

License

This rSkill package (rskill.yaml, README.md, eval/libero.json) is Apache-2.0. The wrapped weights are also Apache-2.0 per the upstream repo. Commercial use is allowed.

OpenRAL
/

rskill-xvla-libero