SmolVLA-SafaiBot-v0 (Phase 0 Stub)

Toy checkpoint from Phase 0 of the safai-vla project.

This is NOT a production model. It is a stub MLP trained on 500 MuJoCo episodes to validate the training pipeline end-to-end. Real model training happens in Phase 1 with Isaac Sim on DGX Cloud.

Architecture

• Base: Stub 4-layer MLP (state -> action), NOT SmolVLA
• Input: 18-dim proprioceptive state (joint_pos + joint_vel + ee_pos + ee_quat + base_pos)
• Output: 7-DOF action
• Parameters: ~105K total
• Training: 20 steps, MSE loss, Adam optimizer

Intended Use

Pipeline validation only. Do not deploy on a real robot.

Source

• GitHub: saralsystems/safai-vla
• Dataset: safai-vla-mujoco-v0
• License: Apache 2.0