rskill-xr1-panda_mobile-robocasa365-nf4

OpenRAL/rskill-xr1-panda_mobile-robocasa365-nf4

rskill-xr1-panda_mobile-robocasa365-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

26% 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-RoboCasa365

Model card

From the Hugging Face model README

XR-1 RoboCasa365

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 RoboCasa365 checkpoint. The package references the pinned upstream checkpoint and preserves its temporal three-camera contract.

Preview

Successful RoboCasa365 OpenDrawer rollout

XR-1 RoboCasa365 OpenDrawer preview

CloseBlenderLid fail rollout

What this skill does

Field	Value
Actions	generalist kitchen manipulation
Objects	kitchen objects and fixtures
Scenes	RoboCasa365 kitchens
Embodiment	`panda_mobile`

How it works

The adapter stores seven observations, samples four frames at offsets [-6, -4, -2, 0], converts the OpenRAL 16-D EE/base quaternion state to the 14-D axis-angle order used by Xiaomi's evaluator, applies the reference 0.95 center crop, and sends the batch to the isolated XR-1 sidecar. It replays 16 actions before replanning.

Observation -> action contract

Direction	Key	Shape	Notes
in	three RGB videos	`(4, H, W, 3)` each	left/right base and wrist
in	state history	`(4, 14)`	EE, gripper, base; raw physical units
out	action chunk	`(T, 12)`	EE delta, gripper, base, padding/mode

Upstream model and training

Field	Value
Source repo	`XiaomiRobotics/Xiaomi-Robotics-1-RoboCasa365`
Base model	XR-1 5B
Paper	Xiaomi-Robotics-1
License	Apache-2.0
Parameters	5.44B
Training data	RoboCasa365

Supported robots

Robot	Embodiment tag	Status	Notes
PandaMobile in RoboCasa	`panda_mobile`	integration-ready	Xiaomi does not pin the exact RoboCasa365 revision, so benchmark reproduction remains blocked.

Sensors required

Three RGB streams plus the PandaMobile EE, gripper, and base state described in the manifest's rc365 bindings.

Manifest summary

Field	Value
`name`	`OpenRAL/rskill-xr1-panda_mobile-robocasa365-nf4`
`version`	`0.1.0`
`license`	`apache-2.0`
`role`	`s1`
`runtime` / dtype	PyTorch sidecar / NF4
`weights_uri`	`hf://OpenRAL/rskill-xr1-panda_mobile-robocasa365-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_robocasa365_close_blender_lid.yaml \
  --rskill rskills/xr1-robocasa365

Reproduction

The scene is a 20-step wiring check matching Xiaomi's documented smoke task, not a score reproduction. Xiaomi's exact RoboCasa365 dependency revision is not published.

Evaluation

No benchmark JSON is shipped until the upstream environment revision can be pinned and the full 50-trial protocol is run.

License

The wrapper and upstream code/weights are Apache-2.0. The checkpoint executes custom HF code and therefore requires OPENRAL_ALLOW_REMOTE_CODE=1.