Event Details
Towards Generalizable Motion Planning: Learning-Based Frameworks for Efficient and Safe Trajectory Generation
Presenter: Mehran Ghafarian Tamizi
Supervisor:
Date: Wed, November 5, 2025
Time: 08:30:00 - 09:15:00
Place: Zoom - see below.
ABSTRACT
Zoom:
Meeting ID: 848 462 6970
Password: 613295
Abstract:
Robotic motion planning remains a fundamental challenge in industrial automation, with manipulators offering a clear example of the need for real-time, collision-free, and safe trajectory generation. Traditional planners often face trade-offs among optimality, adaptability, and computational efficiency, limiting their applicability in cluttered and high-dimensional industrial environments. Furthermore, most learning-based planners suffer from poor generalization, requiring retraining when deployed in new scenes or on different robot platforms. This seminar presents two learning-based frameworks designed to address these challenges. First, we introduce the Path Planning and Collision Checking Network (PPCNet), an end-to-end neural architecture that combines a waypoint generator with a learned collision checker to enable fast, safe, and reliable planning in structured environments. PPCNet is validated in both simulated and real-world bin-picking tasks, demonstrating substantial speed-ups over classical planners while maintaining path quality. To overcome the generalization limitations of PPCNet, we propose Generalizable and Adaptive Diffusion-Guided Environment-aware Trajectory generation (GADGET), a conditional diffusion-based motion planner guided by control barrier functions. GADGET leverages voxel-based scene encoding and goal conditioning to generate safe trajectories across previously unseen environments and robotic arms without retraining. The integration of barrier-function-based guidance ensures robust collision avoidance during trajectory generation. Extensive experiments demonstrate that both frameworks achieve real-time planning performance and high success rates, with GADGET offering strong generalization to novel settings. This work highlights the potential of combining deep generative models with adaptable design to create scalable and broadly generalizable motion planners, capable of transferring across diverse environments and robot platforms with minimal modification.