CoRL2026-CSI/smolvla_IsaacLab-SO101_pick_place_baseCaP_100epi_50ep-appendix

SmolVLA IsaacLab SO101 PickAndPlace 50ep

This repository contains a SmolVLA policy fine-tuned on the IsaacLab SO101 PickAndPlace dataset.

Model

• Policy type: smolvla
• Base model: CoRL2026-CSI/smolVLA-IsaacLab-Multi-Task-8epoch-mod
• VLM backbone: HuggingFaceTB/SmolVLM2-500M-Video-Instruct
• Training mode: frozen vision encoder, expert-only training
• Action chunk size: 50
• Observation steps: 1
• Action dimension: 6
• State dimension: 6

Dataset

• Dataset: CoRL2026-CSI/IsaacLab-SO101-PickAndPlace-100epi-10fps-appendix
• Episodes: 100
• Frames: 34,423
• FPS: 10
• Video codec: H.264
• Cameras used:
  • observation.images.left_wrist renamed to observation.images.camera1
  • observation.images.top renamed to observation.images.camera2

Training

• Steps: 6,724
• Epochs: 50
• Effective global batch size: 256
• Per-GPU batch size: 128
• Number of GPUs: 2
• Optimizer: AdamW
• Peak LR: 1e-4
• Scheduler: cosine decay with warmup
• Decay LR: 2.5e-6
• Seed: 1000
• Image augmentation: color jitter and sharpness jitter
• DataLoader workers: 16 per process
• Preprocessing in workers: enabled

The final checkpoint was saved after step 6724.

Files

• model.safetensors: policy weights
• config.json: policy config
• train_config.json: training config
• policy_preprocessor.json: policy preprocessing pipeline
• policy_postprocessor.json: policy postprocessing pipeline
• policy_preprocessor_step_5_normalizer_processor.safetensors: normalization stats
• policy_postprocessor_step_0_unnormalizer_processor.safetensors: unnormalization stats