Thesis Defense, H.J. Terry Suh. Leveraging Structure for Efficient and Dexterous Contact-Rich Manipulation
Thesis Defense: Leveraging Structure for Efficient and Dexterous Contact-Rich Manipulation
Speaker: H.J. Terry Suh
In-person Location: 32-D463 (Star)
Zoom Link: https://mit.zoom.us/j/7652207066
Abstract:
Contact-Rich Manipulation has proved challenging due to the need to consider multiple combinatoric possibilities of making or breaking contact with the surrounding environment. As a result, existing methods have often resorted to combinatorial optimization that utilizes dynamics structure but considers all possibilities exhaustively, or compute-heavy and inefficient sampling methods that utilize blackbox optimization such as Reinforcement Learning (RL). In this thesis, I aim to show that by combining structured contact smoothing in conjunction with local gradient-based control and sampling-based motion planning, we can bypass the combinatorial explosion of contact modes while still leveraging structure and achieving highly efficient contact-rich manipulation. Foremost, I shed light on how RL abstracts contact modes and optimizes difficult landscapes by combining stochastic smoothing and zeroth-order optimization; yet, I show how following a similar stochastic strategy while using gradients suffers from several drawbacks such as empirical bias and high variance. To leverage structure in a more helpful manner, I propose a method for smoothing contact dynamics without relying on stochastic smoothing, bypassing these drawbacks. Using this smoothing scheme, I show that we can achieve highly efficient and performant local control using gradient-based trajectory optimization and model predictive control. Finally, I remark on some fundamental limitations of local control and how we can connect the proposed local control capability with more global sampling-based motion planners to achieve long-horizon global plans. The proposed method achieves contact-rich plans such as dexterous in-hand reorientation and whole-body manipulation much more efficiently than RL while being highly scalable compared to methods that explicitly reason about contact modes. These results achieve a reduction of contact-rich manipulation to kinodynamic motion planning, and exposes the true difficulty of contact-rich manipulation from combinatorial explosion in contact modes to combinatorial and highly non-local decisions over motion planning behaviors.
Speaker: H.J. Terry Suh
In-person Location: 32-D463 (Star)
Zoom Link: https://mit.zoom.us/j/7652207066
Abstract:
Contact-Rich Manipulation has proved challenging due to the need to consider multiple combinatoric possibilities of making or breaking contact with the surrounding environment. As a result, existing methods have often resorted to combinatorial optimization that utilizes dynamics structure but considers all possibilities exhaustively, or compute-heavy and inefficient sampling methods that utilize blackbox optimization such as Reinforcement Learning (RL). In this thesis, I aim to show that by combining structured contact smoothing in conjunction with local gradient-based control and sampling-based motion planning, we can bypass the combinatorial explosion of contact modes while still leveraging structure and achieving highly efficient contact-rich manipulation. Foremost, I shed light on how RL abstracts contact modes and optimizes difficult landscapes by combining stochastic smoothing and zeroth-order optimization; yet, I show how following a similar stochastic strategy while using gradients suffers from several drawbacks such as empirical bias and high variance. To leverage structure in a more helpful manner, I propose a method for smoothing contact dynamics without relying on stochastic smoothing, bypassing these drawbacks. Using this smoothing scheme, I show that we can achieve highly efficient and performant local control using gradient-based trajectory optimization and model predictive control. Finally, I remark on some fundamental limitations of local control and how we can connect the proposed local control capability with more global sampling-based motion planners to achieve long-horizon global plans. The proposed method achieves contact-rich plans such as dexterous in-hand reorientation and whole-body manipulation much more efficiently than RL while being highly scalable compared to methods that explicitly reason about contact modes. These results achieve a reduction of contact-rich manipulation to kinodynamic motion planning, and exposes the true difficulty of contact-rich manipulation from combinatorial explosion in contact modes to combinatorial and highly non-local decisions over motion planning behaviors.