mateuszwasko1/smolvla_dual_arm_m3

SmolVLA Dual-Arm (Multi-Object, Full Domain Randomization)

This model is a fine-tuned SmolVLA (Small Vision-Language-Action) policy trained for cooperative dual-arm manipulation in PyBullet. It is designed to handle multiple diverse objects under heavy visual domain randomization.

Model Details

• Architecture: SmolVLA + PEFT LoRA (Rank 64).
• Task: Dual-arm cooperative lifting and placing of various objects.
• Action Space: 12-D absolute joint angles in radians (6 degrees of freedom per arm, gripper state fixed).
• Vision: 3 camera streams (overhead, left wrist, right wrist) at 224x224 resolution.
• Training Data: 140 expert demonstrations collected at 20 FPS with full domain randomization applied at collection time (randomized table/floor colors, lighting conditions, object textures, and scale ±15%).

Performance

By leveraging the pretrained vision-language backbone and domain-randomized data, this policy demonstrates strong zero-shot visual robustness. It successfully generalizes across a variety of unseen visual conditions, completely eliminating the visual degradation observed in non-randomized baselines.