vla-mjlab-piper-stack-act-50k

ACT Piper Stack v4 Chunk100 50K

Best checkpoint from the vla-mjlab Piper stacking ACT run.

Result

Evaluated headlessly in MJLab with 100 rollout trials:

• Ever success: 81/100 (81%)
• Stable after 1.5s: 76/100 (76%)
• Final success: 76/100 (76%)

This checkpoint outperformed later checkpoints despite higher training loss, so use this 50K checkpoint for the current ACT baseline.

Dataset

Trained on axiboai/vla-mjlab-piper-stack-act, success-only episodes from the piper_stack_act_v4 dataset.

Schema:

• 3 cameras: wrist left, wrist right, scene top
• observation.state: 7D right arm qpos + measured gripper
• action: 7D right arm target qpos + target gripper
• Control rate: 50 Hz

Policy

LeRobot ACT configuration:

• chunk_size=100
• n_action_steps=100
• n_decoder_layers=7
• n_encoder_layers=4
• dim_model=512
• dim_feedforward=3200
• kl_weight=10
• batch_size=8
• trained checkpoint step: 50K

Rollout

Use the sagar-dev branch of axiboai/vla-mjlab:

python -m scripts.play_act_viser \
  --policy-path axiboai/vla-mjlab-piper-stack-act-50k \
  --device cuda \
  --num-envs 4 \
  --chunk-consume 100 \
  --freeze-on-success \
  --freeze-success-delay-steps 75

The local rollout script applies the saved LeRobot preprocessors and postprocessors, and builds viewer cameras at the checkpoint's expected image size.