rskill-smolvla-franka_panda-vlabench-bf16

rskill-smolvla-franka_panda-vlabench-bf16

OpenRAL rSkill — VLABench integration baseline (NOT a passing policy). SmolVLA (~0.5 B) finetuned on VLABench (lerobot/vlabench_unified), lerobot-native, runs in-process on lerobot 0.6.0 (bf16, fits 8 GB). Wraps lerobot/smolvla_vlabench.

⚠ Status: baseline, scores 0%

This rSkill exists to exercise and validate the OpenRAL VLABench backend, not to score the benchmark. Measured 0/3 on six diverse primitive tasks (select_fruit, select_drink, select_toy, select_book, add_condiment, insert_flower) — identical to lerobot's own lerobot-eval reference, which confirms the OpenRAL wiring is faithful (state 7-D, action absolute-eef, cameras camera1/2/3) and the 0% is the policy, not the integration.

The only VLABench policy above 50% is VLABench/pi0-fast-ft-primitive-10task-deltachunk (51.2% primitive avg), which is openpi/JAX and would need conversion to lerobot PI0FAST (+int8 for 8 GB) to run in-process — a dedicated, not-yet-done effort. VLABench's composite/long-horizon suite is unsolved (<50%) by every known policy.

Provisioning

The Python side auto-installs on first env build via the vlabench ensure_backend_deps plan (OPENRAL_AUTO_INSTALL_DEPS=1, the default): it clones OpenMOSS/VLABench, uv pip install --no-deps -e's it, adds the numpy-2 sim deps, and writes a raise-on-use rrt_algorithms stub (git-only data-gen dep, off the VLA eval path).

The ~12 GB CC-BY asset bundle is a one-time manual fetch (a Google-Drive gdown pull too flaky to drive unattended — the backend raises with this exact recipe when it is absent):

export VLABENCH_ROOT=$HOME/.cache/openral/repos/VLABench/VLABench   # the clone the plan installs
python $HOME/.cache/openral/repos/VLABench/scripts/download_assets.py   # ~12 GB obj + scene

Run

MUJOCO_GL=egl VLABENCH_ROOT=$VLABENCH_ROOT \
  openral benchmark scene --config scenes/benchmark/vlabench_select_fruit.yaml \
    --rskill rskills/smolvla-vlabench

Upstream model / training

Base is lerobot/smolvla_base (SmolVLA ~0.5 B, arXiv:2506.01844), finetuned on lerobot/vlabench_unified (VLABench, 97 tasks). The wrapped checkpoint is lerobot/smolvla_vlabench (Apache-2.0); OpenRAL adds no weights, only packaging.

Supported robots / embodiments

franka_panda (VLABench's 7-DOF Franka Panda). The manifest's embodiment_tags must intersect the robot's — matched against robots/franka_panda.

Sensors / observation contract

Three RGB views (camera1/2/3, from the env's image/second_image/wrist_image, ≥224×224) plus a 7-D proprio state [pos_robot(3), euler_xyz(3), gripper(1)]. The checkpoint's preprocessor renames image→camera1… (a no-op on the already-canonical keys) and resizes to 256.

Manifest summary

License

Apache-2.0 — both this rSkill's packaging and the wrapped lerobot/smolvla_vlabench checkpoint.

See also

• python/sim/src/openral_sim/backends/vlabench.py — the backend.
• lerobot/smolvla_vlabench — upstream checkpoint.