Teaching robots to manipulate objects with reinforcement learning
A reinforcement learning framework for training robotic arms to perform complex manipulation tasks like stacking, sorting, and assembly in simulation and real-world settings.
Programming robots for manipulation tasks is time-consuming and brittle. Traditional approaches require precise specifications and fail when encountering variations in object positions or unexpected obstacles.
Developed a sim-to-real transfer pipeline using domain randomization and asymmetric actor-critic methods. Implemented curriculum learning strategies that progressively increase task difficulty. Created a modular reward shaping framework for rapid task specification.
Achieved 92% success rate on real-world pick-and-place tasks after training purely in simulation. Reduced robot programming time from weeks to hours for new manipulation tasks.