Every Embodied SmolVLA MuJoCo Pick-and-Place

This is the Datawhale staging repository for the SmolVLA checkpoint used in the Every Embodied AMD ROCm policy reproduction course.

Current status

The protected rebuilt checkpoint in weights/ is the weighted_000500 checkpoint. It passed the strict closed-loop protocol with:

• red-mug instruction: 27/30 strict physical successes;
• blue-mug instruction: 30/30 strict physical successes;
• overall: 57/60 (95.00%).

This is a fine-tuned SmolVLA policy, not a zero-shot foundation-model result. The matching dataset manifest, normalization/preprocessing configuration and evaluation summaries are intentionally documented separately from the model file. Raw training data and optimizer state are not included.

Task and interface

• Simulator: MuJoCo SimpleEnv2
• Robot: OMY-style arm
• Observations: two RGB cameras, 6-D robot state, and a language instruction
• Action: 7-D joint/gripper command
• Control frequency: 20 Hz
• Instructions:
  • Place the red mug on the plate.
  • Place the blue mug on the plate.

Strict evaluation

The preserved evaluation protocol uses forced red/blue instructions, seeds 0-29, and at most 600 action steps. Physical success requires the legacy task predicate plus a real lift of at least 0.03 m for at least three control ticks and a final upright cosine of at least 0.7.

• Strict evaluation summary
• Representative blue-mug success video

The protected model file is:

• weights/model.safetensors
• SHA256: abc335dbd4d4fdbfca8b188b22588b3d768747c9b4ee3051593a4a506c893e0d

The complete file list and checksums are in SHA256SUMS.txt.

Loading the protected checkpoint

Download this repository with huggingface_hub or the Hugging Face web UI, then point the Every Embodied SmolVLA Notebook at the local weights/ directory. The model must be evaluated with the matching camera order, normalization files, action bridge and strict physical-success predicate. It should not be evaluated with the ACT/Pi0 action interface.

Artifact scope

This repository contains the protected model artifact, its configs, checksums and strict evaluation evidence. The raw dataset and optimizer state are omitted; use the course's documented data preparation and training Notebook to reproduce the recipe.

Course repository: datawhalechina/every-embodied