All projectsRobot learning, VLA
Language-conditioned manipulation
SmolVLA pick-and-place on dual Franka arms
SmolVLA as the policy for bimanual pick-and-place on two Franka arms in Genesis. It takes several camera views and a typed instruction and outputs one action for both arms.
I evaluated completion rate, grasp success and failure modes under randomised lighting and object placement, and used the breakdown to decide what to fix each round.
- Type
- Simulation
- Period
- 2026
- Role
- ML engineering, training and evaluation
- Stack
- SmolVLA, LeRobot, Genesis, Franka Emika Panda, Python, PyTorch
- Result
- Colour sort: 6 of 6 cubes in 57 s. Language-targeted task: 4 of 4 targeted cubes, distractors ignored.
What I built
- Language-conditioned bimanual manipulation
- Multi-camera input to a single VLA policy
- Evaluation under domain randomisation
- Each training round driven by the failure breakdown