Geometric Fabrics: Generalizing Classical Mechanics to Capture the Physics of Behavior

Abstract

Fig. 1. Reinforcement learning over a geometric fabric layer yields safe, high-performance manipulation behavior for a highly-actuated hand. The learned behavior switches between two-, three-, and four-fingered grasps during prehensile manipulation. Videos at https://dextreme.org/fgp.html.

Dynamic Optimization Fabrics for Motion Generation

“…Experiments in dynamic environments showed that DF increase clearance to moving obstacles without significantly increasing path length, execution time and computational costs. Thus, DF overcome an important drawback of SF [9,10], where collision avoidance with moving obstacle is solved purely by the high frequency at which optimization fabrics can be computed. Moreover, the generalization did not increase the solving time compared to SF.…”

“…While the previous subsection defined what criteria are required for a spec to form an optimization fabrics, the theory on conservative fabrics and energization offers a simply way of generating such special specs. As a full summary of the theory on optimization fabrics and their construction is out of scope here, this subsection only provides an outline of the theory and the reader is referred to [28,10] for detailed derivations.…”

“…We compared an MPC planner with a static fabric planner [9,10]. The low-level controller of the robot runs at 1 kHz.…”

“…This allows to iteratively design motion while maintaining explainability over the resulting motion. Despite these advantages, optimization fabrics [8,9,6,10] do not integrate the dynamics of the environment. Besides, as a highly reactive method, optimization fabrics are prone to local minima.…”

DexTransfer: Real World Multi-fingered Dexterous Grasping with Minimal Human Demonstrations

“…Geometric fabrics is a provably stable, second order policy that effectively solves the problem of reaching to end-effector targets while resolving arm redundancy, controlling manipulator posture, avoiding joint limits and joint speed limits, avoiding robot self-collision, and avoiding excessive collision between the robot and the table. The design and tuning of this policy is exactly the same as the one reported in [41].…”