Shiki42/s016-sortblocks-sequential-act-step100000

S016 Blocks Ranking Sequential ACT, step 100000

Official LeRobot 0.4.4 ACT policy trained for the RoboTwin Blocks Ranking RGB-CTR simulation task. This repository contains the preselected final inference checkpoint from CTR Experiment E766, Run E766-R001.

Training identity

• Training dataset: Shiki42/ctr-sortblocks-100ep-sequential at immutable revision 3d8db515f856c4e6c2612569b43fb15eeaef3064 (100 episodes, 91,191 frames).
• ACT training: 100,000 optimizer updates, effective batch size 8, seed 87431, S016 Sequential arm with 50 left-first and 50 right-first whole episodes.
• Source: official LeRobot 0.4.4 commit 8fff0fde7c79f23a93d845d1a50e985de01f8b8a; CTR launch code commit 4d3aabc65e93878f7c9bf1b8a5900d35fd6bcf2e.
• Non-CTR dataset: IdleMask disabled. Saved LeRobot preprocessor and postprocessor provide the training normalization and inverse transform.
• Cameras: top and centered fovy90 left/right wrist RGB; state and action vectors have 14 dimensions.

Checkpoint and status

The formal Run completed with exit code 0 on 2026-09-23 after all 100,000 updates. The final checkpoint passed a fresh-process CPU reload: 234 finite model state tensors matched model.safetensors exactly, and both saved processors loaded. The model file SHA-256 is 2c4df9c87e941644b9b15d29fce0d4c5663f741f456bbbe9180dd11d6a4c0633. The complete public inference package also includes SHA256SUMS.

Held-out RoboTwin success and collision rates have not been established. The shared held-out 100-scene suite and ACT evaluation adapter are still awaiting qualification (evaluation Experiment E781). Do not treat training completion as a task-success result.

Optimizer and RNG training state are retained in the external formal Run artifacts and are not part of this inference repository.