Generative Motion Planning for Robotics

A robotics workflow for predicting and optimizing motion through dynamic, cluttered environments.

Reliable robot manipulation requires perception and learned predictions to connect cleanly with kinematic and dynamic constraints.

Calibrated an OAK-D Pro stereo camera, developed PyTorch-based generative models, and built the Python trajectory optimization pipeline in Drake's digital twin environment.

Stereo observations feed perception and a learned model that proposes candidate trajectories before constrained optimization produces robot motion.

1. 01

   Learned trajectory proposals refined by constrained optimization

   Instead of

   An end-to-end learned control policy

   Why

   An end-to-end policy can produce motion that violates joint limits or dynamics constraints with no guarantee of feasibility. Using the model to propose good initial guesses and an optimizer to enforce constraints keeps the speed benefit of learning while keeping every output physically valid.

2. 02

   Drake for modeling and trajectory optimization

   Instead of

   MoveIt with sampling-based planners

   Why

   Sampling-based planners handle collision-free routing well but treat dynamics as an afterthought. Drake's multibody model and differentiable optimization made it possible to optimize directly against kinematic and dynamic constraints, which is what the research question required.

• Python
• PyTorch
• Drake
• OAK-D Pro
• Computer Vision