Work / Robot learning, VLA Simulation

Language-conditioned manipulation

SmolVLA pick-and-place on dual Franka arms

A typed instruction selects the action; both arms execute it.Simulation

This pipeline uses SmolVLA as the policy for bimanual pick-and-place on dual Franka arms in the Genesis simulator. SmolVLA takes multiple camera views and a language instruction and outputs a combined action for both arms.

Evaluation focused on task completion rate, grasp success and failure-mode breakdowns under randomised lighting and object placement, and these results drove each round of policy iteration.

In brief

Language-conditioned bimanual manipulation
Multi-camera input to a single VLA policy
Evaluation under domain randomisation
Data-driven iteration from failure analysis

From the project

Colour sorting: a vision-language model plans which cube goes where and inverse kinematics moves the arms.Simulation