OpenRAL/rskill-xr1-panda_mobile-robocasa-nf4
XR-1 RoboCasa
OpenRAL wrapper for Xiaomi Robotics' XR-1 RoboCasa checkpoint on the PandaMobile kitchen embodiment. The package points at the pinned upstream safetensors; it does not copy weights.
Preview

What this skill does
How it works
XR-1 combines Qwen3-VL-4B-Instruct with a 36-layer DiT action head. OpenRAL runs the exact torch 2.8 / transformers 4.57.1 / FlashAttention stack in a sidecar, pads the 8-D physical state to XR-1's internal 60-D state tensor, and applies Xiaomi's 0.95 center crop before returning the checkpoint's first seven decoded action channels.
Observation -> action contract
Upstream model and training
Supported robots
Sensors required
Three RGB streams at 256x256 or larger and an 8-D arm/gripper state.
Manifest summary
Quick start
just sync --group sidecar-wire --group robocasa
OPENRAL_ALLOW_REMOTE_CODE=1 openral sim run \
--config scenes/sim/xr1_robocasa_pnp.yaml \
--rskill rskills/xr1-robocasaReproduction
The command above validates the OpenRAL observation/action wiring. It is not a reproduction of Xiaomi's published RoboCasa score because Xiaomi evaluates RoboCasa v0.2 while OpenRAL currently provisions its own pinned kitchen stack.
Evaluation
No benchmark JSON is shipped until a full upstream-version-matched run is completed. Integration claims are limited to schema and adapter validation.
License
The rSkill metadata and upstream XR-1 code/checkpoint are Apache-2.0. Loading requires OPENRAL_ALLOW_REMOTE_CODE=1 because the HF repo ships executable custom Transformers code.
See also
- `scenes/sim/xr1_robocasa_pnp.yaml`
- `docs/reference/vla_compatibility.md`
