VLA-Adapter SO-101 Push Cube — Step 10,000
This repository contains an intermediate 10,000-optimizer-step checkpoint of
VLA-Adapter fine-tuned for the SO-101 push_cube task.
Training configuration
| Setting | Value |
|---|---|
| Dataset | taewonkoo/push_cube_v2_rlds |
| Optimizer step | 10,000 / 20,000 |
| Per-device batch size | 4 |
| Gradient accumulation | 4 |
| Effective batch size | 16 |
| Learning rate | 1e-4 |
| Scheduler | Cosine with 2,000 warmup steps |
| LoRA rank | 32 |
| Input images | 2 (front, top) |
| Proprioception | Enabled, 6 dimensions |
| Action dimension | 6 |
| Action chunk length | 8 |
| Precision | BF16 on an RTX 3090 |
The RLDS dataset contains 50 trajectories and 13,679 transitions. Actions and proprioceptive states follow the six-motor SO-101 representation.
Checkpoint contents
model.safetensors: VLA model with the LoRA weights mergedlora_adapter/: trainable LoRA adapteraction_head--10000_checkpoint.pt: continuous action headproprio_projector--10000_checkpoint.pt: proprioception projectorvla_extras--10000_checkpoint.pt: trainable VLA extras such as action queriestraining_state--10000_checkpoint.pt: optimizer, scheduler, scaler, and RNG state- processor, tokenizer, configuration, and dataset statistics
The separate action head and proprio projector are required for the SO-101
continuous-control policy. The merged model.safetensors alone is not the
complete policy.
Code
Use the SO-101-compatible training and inference code from:
https://github.com/TaewonKoo/VLA-Adapter-SO101
This is an intermediate training checkpoint, not the final 20,000-step model. Review the base model, dataset, and code licenses before redistribution or deployment. Validate the policy in simulation or a controlled robot workspace before operating physical hardware.