physics-reasoning-vla checkpoints
Final checkpoints (47,064 steps, batch 24) of SmolVLM2-500M + SmolVLA action-expert policies trained on a simulated manipulation
dataset with four tasks (pick, bounce, basket, curling). Each folder is a LeRobot pretrained_model directory (config.json,
model.safetensors, pre/post-processor files). Loading requires the lerobot_policy_smolvla_phys plugin from the project code
repository; dataset-relative paths inside config.json are written as ${DATA_ROOT} and must be pointed at the dataset root.
| folder | reasoning tokens | track token | property head | FiLM |
|---|---|---|---|---|
| plain1f_none (1 frame per camera) | - | - | - | - |
| plain3f_none | - | - | - | - |
| baseline_tokens | x | - | - | - |
| trackonly_none_extdet (external detector) | - | x | - | - |
| lochead_tokens_notrack / lochead_none_notrack (localisation head only) | x / - | - | - | - |
| R6cell_none_track_head | - | x | x | - |
| R6_tokens_track_head | x | x | x | - |
| R8a_tokens_track_head_timing | x | x | x (+timing targets) | - |
| R8bnone_none_track_head_timing_film | - | x | x (+timing) | x |
| R8b_tokens_track_head_timing_film | x | x | x (+timing) | x |
| film2_none | - | x | x | x |
source_run_name.txt in each folder gives the original training run name.