ACT vs SmolVLA: Testing on Hugging Face’s LeRobot SO-101

This video compares two robot control policies — Action Chunking Transformer (ACT) and SmolVLA — using the LeRobot SO-101 setup from Hugging Face. ACT: An imitation learning policy that directly maps visual input to joint commands. It performs well but can be overly aggressive when correcting anomalies. SmolVLA: A smaller Visual-Language-Action (VLA) foundation model. It performs more deliberate and stable motions, reattempting tasks when it conceptually understands “picking.” SmolVLA also supports language-conditioned instructions, enabling the robot to perform tasks described in natural language (e.g., “pick up the blue bottle” or “move the box to the right”). However, this language capability is not showcased in this video. Both models are tested under similar conditions to showcase differences in robustness, behavior, and stability.