HuggingFace LeRobot SmolVLA fine-tuned on the ManiGuard clutter base task (sim Franka Panda). Part of the ManiGuard VLA benchmark - SmolVLA vs pi0.5 vs GR00T on the same task families with identical data, cameras, and controller.

Model

Base: lerobot/smolvla_base - SmolVLM2 vision-language backbone + flow-matching action expert
Embodiment: Franka Panda, 8-D joint state/action (7 arm joints + 1 gripper), padded to SmolVLA's 32-D
Cameras (2): observation.images.top (overview = left view) + observation.images.wrist (256x256)
Action: absolute joint targets (fed straight to a JointController at eval); 50-step chunk
Tuning: SmolVLA default - vision encoder frozen, train the action expert (no LoRA)

Training

8-GPU accelerate DDP, global batch 512, 3522 steps (~2 epochs over 901,520 frames)
Data: IDEAS-Lab-Northwestern/datagen-clutter-v1-joint-5cam - the 5-cam datagen dataset, prepared to a 2-cam standard-keyed copy (H.264 videos)

Usage

Load with SmolVLAPolicy.from_pretrained("IDEAS-Lab-Northwestern/smolvla-base-datagen-v1-clutter-joint-2cam-promptltl-yanZ") from LeRobot. The checkpoint carries the normalization stats.

WARNING - Convention (must match at eval): joint-space JointController (absolute joint targets) + 2 cameras (observation.images.top = the left overview, observation.images.wrist). A mismatched controller, camera set, or overview view silently feeds an out-of-distribution input.