SO-101 Stack3 ฯ€0.5 baseline checkpoints

This public repository contains five deployment snapshots from one Stack3 fine-tuning run. Checkpoints are stored by optimizer step:

Each checkpoint directory contains the model weights, model configuration, tokenizer, and matching observation/action pre- and post-processors. The optimizer and training-resume state are not included in these deployment snapshots.

Training setup

  • Robot: SO-101, single arm; front and wrist RGB cameras
  • Observation state: 6 dimensions
  • Action: 6 dimensions; first five arm-relative, gripper absolute
  • Chunk size: 50; execution horizon: 10 actions
  • Precision: BF16; batch size: 64
  • Initialization: Stack2 ฯ€0.5 checkpoint 004228, followed by Stack3 full-chain fine-tuning
  • Dataset: 64 Stack3 full-chain episodes (60,481 frames), all used for training; no held-out validation split
  • Fine-tune learning rate: peak 1.25e-5, cosine decay to 1.25e-6, 1,000-step warmup
  • Training completed: 10,000 optimizer updates

These are deployment candidates, not a claim of successful three-block stacking. The checkpoints have not been validated on the physical robot. Training loss does not establish real-world task success.

The Stack2 initialization was SO-101 Stack2 ฯ€0.5, checkpoint 004228. No OCBench, FurnitureBench, or OGBench trajectories were used. The Stack3 dataset and training environment are private and are not included in this repository.

Downloads last month

-

Downloads are not tracked for this model. How to track
Video Preview
loading