lerobot / being_h05_robocasa

Being-H0.5 RoboCasa — LeRobot conversion

This is a deterministic LeRobot-format conversion of BeingBeyond/Being-H05-2B_robocasa . It contains the policy configuration, serialized preprocessor and postprocessor pipelines, tokenizer, conversion report, and safetensors weights.

Revisions

• Author source: ab1204954cb5e68f0bc52a376982e962c763bbbe
• Source checkpoint: 75989b121eab7d0a8e421d846acf268c2938883d
• LeRobot integration: 44c942c9bf648ea27729cae1b38eb8d33377652e

Validation

• 990 source tensors mapped to 1,086 native-layout tensors
• Transformers-native Qwen3Config and InternVLVisionModel
• 0 missing, unexpected, duplicate, or shape-mismatched tensors
• Strict offline save/reload passed
• SHA-256: bbe5c8d264e705296851b846e625e936fb1bf05f92b39d311777cb59d0e57aa1
• Author-vs-LeRobot preprocessing, flow velocity/loss, sampled action, and inverse action projection: maximum and p99 absolute error 0
• 50 RoboCasa episodes: author 12/50 , LeRobot 12/50 , delta 0 percentage points, Wilson 95% CI 14.3%–37.4%
• 3,892 online policy requests: maximum and p99 action error 0

The checkpoint preserves Being-H0.5's fixed semantic 200D state/action slots, validity masks, three-camera processing, raw task string, and four-step flow-matching sampler.

Usage

The audited author BeingH/MoT action architecture is currently required at runtime; the Qwen configuration, tokenizer, and InternVL vision tower are provided by Transformers:

git clone https://github.com/BeingBeyond/Being-H.git
git -C Being-H checkout ab1204954cb5e68f0bc52a376982e962c763bbbe
export PYTHONPATH="$PWD/Being-H/Being-H05:$PYTHONPATH"

Then load the checkpoint as a standard LeRobot policy:

lerobot-eval --policy.path=lerobot/being_h05_robocasa