rskill-xr1-panda_mobile-robocasa-nf4

OpenRAL/rskill-xr1-panda_mobile-robocasa-nf4

rskill-xr1-panda_mobile-robocasa-nf4 is a robotics model from OpenRAL. Use it for the robotics task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.

Requires OpenRAL Pro. This rSkill runs at int4 (NF4) (sub-16-bit weights). Under ADR-0123, open-source OpenRAL loads only fp32, fp16 and bf16 weights; quantized formats (int8, fp8, int4/NF4, nvfp4) need the OpenRAL Pr…

Downloads · 30 days

38% of all-time downloads

All-time downloads

Public

Parameters

5.1B

3.2 GB on disk

Likes

Public

Hugging Face

Repo makeup

Click a slice to open those files.

.safetensors3.2 GB · 99%

Parameter types

How the weights are stored.

U84.7B · 92%

Base models

XiaomiRobotics/Xiaomi-Robotics-1-RoboCasa

Model card

From the Hugging Face model README

XR-1 RoboCasa

Requires OpenRAL Pro. This rSkill runs at int4 (NF4) (sub-16-bit weights). Under ADR-0123, open-source OpenRAL loads only fp32, fp16 and bf16 weights; quantized formats (int8, fp8, int4/NF4, nvfp4) need the OpenRAL Pro quantization package.

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

Clean RoboCasa rollout (fail)

XR-1 RoboCasa preview

What this skill does

Field	Value
Actions	pick, place, open, close
Objects	kitchen objects
Scenes	RoboCasa kitchens
Embodiment	`panda_mobile`

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

Direction	Key	Shape	Notes
in	`camera1`, `camera2`, `camera3`	three RGB 256x256 frames	left/right base and wrist
in	`observation.state`	`(8,)`	seven arm joints plus one gripper value
out	action chunk	`(T, 7)`	delta XYZ, delta axis-angle, gripper

Upstream model and training

Field	Value
Source repo	`XiaomiRobotics/Xiaomi-Robotics-1-RoboCasa`
Base model	XR-1 5B post-trained for RoboCasa
Paper	Xiaomi-Robotics-1
License	Apache-2.0
Parameters	5.44B
Training data	Xiaomi's RoboCasa benchmark post-training data

Supported robots

Robot	Embodiment tag	Status	Notes
PandaMobile in RoboCasa	`panda_mobile`	integration-ready	Upstream evaluates RoboCasa v0.2; OpenRAL's current backend version must still be benchmark-correlated.

Sensors required

Three RGB streams at 256x256 or larger and an 8-D arm/gripper state.

Manifest summary

Field	Value
`name`	`OpenRAL/rskill-xr1-panda_mobile-robocasa-nf4`
`version`	`0.1.0`
`license`	`apache-2.0`
`role`	`s1`
`runtime` / dtype	PyTorch sidecar / NF4
`weights_uri`	`hf://OpenRAL/rskill-xr1-panda_mobile-robocasa-nf4`
`latency_budget.per_chunk_ms`	120000
`commercial_use_allowed`	yes

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-robocasa

Reproduction

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.