IPAB Workshop - 05/06/2025 Speaker: Haocheng Yuan Title: Neural Computer-Aided-Design Generation Abstract: Deep learning-based generative models are now capable of automatically producing complex, editable 3D CAD models, transforming the design process across engineering and architectural domains. Key generative modeling trends driving these advances include autoregressive modeling of CAD construction sequences (treating design operations as sequential commands), diffusion models that generate 2D and 3D geometry, and structured latent representations that capture the underlying parametric and topological constraints of designs. These techniques have been applied to a wide range of CAD domains, including mechanical part generation, automated architectural floorplan layout synthesis, and parametric sketch generation for early-stage design ideation. The research background and several recent advances will be included during the workshop. Speaker: Victor Leve Usage Leve-Lin Title: Scaling Whole-Body Manipulation with Contact Optimization Abstract: Daily tasks require us to use our whole body to manipulate objects, for instance when our hands are unavailable. We consider the issue of providing humanoid robots with the ability to autonomously perform similar whole-body manipulation tasks. Ideally, a human should only need to specify the desired pose of the object to manipulate, while the corresponding sequence of contacts and joint trajectory to accomplish the task is computed autonomously. However, the infinite possibilities for where and how contact can occur on the robot and object surfaces hinder the scalability of existing planning methods, which predominantly rely on discrete sampling. The continuous nature of contact surfaces makes continuous optimization of contacts a natural choice for guiding the planning to converge faster to a solution. Yet, to enable this two challenges remain: (i) Finding a differentiable representation of the non-convex surfaces of the robot and the object; (ii) Finding a strategy for deciding autonomously where and how to contact between these surfaces. In our work we study novel continuous representations of the robot surface and decision strategies for enabling whole-body manipulation planning with continuous optimization of contacts. We demonstrate the improvement of planning performance offered by our approach compared to the state of the art, as well as its suitability for hardware transfer. Jun 05 2025 13.00 - 14.00 IPAB Workshop - 05/06/2025 Haocheng Yuan & Victor Leve Usage Leve-Lin MF2
IPAB Workshop - 05/06/2025 Speaker: Haocheng Yuan Title: Neural Computer-Aided-Design Generation Abstract: Deep learning-based generative models are now capable of automatically producing complex, editable 3D CAD models, transforming the design process across engineering and architectural domains. Key generative modeling trends driving these advances include autoregressive modeling of CAD construction sequences (treating design operations as sequential commands), diffusion models that generate 2D and 3D geometry, and structured latent representations that capture the underlying parametric and topological constraints of designs. These techniques have been applied to a wide range of CAD domains, including mechanical part generation, automated architectural floorplan layout synthesis, and parametric sketch generation for early-stage design ideation. The research background and several recent advances will be included during the workshop. Speaker: Victor Leve Usage Leve-Lin Title: Scaling Whole-Body Manipulation with Contact Optimization Abstract: Daily tasks require us to use our whole body to manipulate objects, for instance when our hands are unavailable. We consider the issue of providing humanoid robots with the ability to autonomously perform similar whole-body manipulation tasks. Ideally, a human should only need to specify the desired pose of the object to manipulate, while the corresponding sequence of contacts and joint trajectory to accomplish the task is computed autonomously. However, the infinite possibilities for where and how contact can occur on the robot and object surfaces hinder the scalability of existing planning methods, which predominantly rely on discrete sampling. The continuous nature of contact surfaces makes continuous optimization of contacts a natural choice for guiding the planning to converge faster to a solution. Yet, to enable this two challenges remain: (i) Finding a differentiable representation of the non-convex surfaces of the robot and the object; (ii) Finding a strategy for deciding autonomously where and how to contact between these surfaces. In our work we study novel continuous representations of the robot surface and decision strategies for enabling whole-body manipulation planning with continuous optimization of contacts. We demonstrate the improvement of planning performance offered by our approach compared to the state of the art, as well as its suitability for hardware transfer. Jun 05 2025 13.00 - 14.00 IPAB Workshop - 05/06/2025 Haocheng Yuan & Victor Leve Usage Leve-Lin MF2
