MARS: Tube Transport checkpoints

Pretrained weights for the MARS Panda tube-transfer simulation. Source code and installation: SkylineZ123/MARS.

Files

Path	Contents
mars/mars.pt	One MARS router checkpoint; candidate order: vla, pickup, insert
actors/pickup.pt	One pickup expert Actor checkpoint
actors/insert.pt	One insertion expert Actor checkpoint
adapter/adapter.pt	Shared RL Token adapter
vla/	Frozen fine-tuned SmolVLA, tokenizer, preprocessing and normalization resources
simulation.yaml	Relative paths and execution settings
manifest.json	SHA256 file-integrity manifest

The router and both experts belong to the same training run. Expert checkpoints contain inference Actors, without critics, optimizers or replay. The VLA executes 10 actions per query; each expert produces 5 actions.

Run

Install the MARS source and its simulation dependencies as documented in the source repository. From that checkout, using this downloaded folder as assets/tubes:

export NUMBA_DISABLE_JIT=1
export MUJOCO_GL=egl
mars-assets verify --assets assets/tubes
mars-tubes run --assets assets/tubes --seed 0

For optional evaluation, select the only provided model seed explicitly:

mars-tubes evaluate --assets assets/tubes --model-seeds 0 --modes mars

Dependencies and installation instructions are maintained with the source code, not bundled in this weights repository. Component attribution is retained in attribution/.