MIT FutureTech · MIT CSAIL
VLA Manipulation Research
An end-to-end MolmoAct2 scaling and evaluation pipeline for Franka manipulation, spanning data generation, fine-tuning, custom tasks, and policy evaluation.
Franka task reelPrimary VLA video slot
Peg-in-hole taskVideo slot
Evaluation viewVideo or figure slot
Measure what
scaling changes.
I built the research pipeline in MolmoSpaces and MuJoCo with LeRobot and MolmoAct2 VLA policies. Custom peg-in-hole tasks enabled controlled evaluation across task precision and data volume.
The pipeline connected successful expert demonstrations, LoRA fine-tuning, and rollout evaluation into one repeatable workflow.
MuJoCoSimulation
LeRobotData and policy tooling
LoRAFine-tuning