Contact-Implicit Stein Projected ADMM for Discovery of Diverse Contact-Rich Manipulation Strategies

Anonymous Authors
Paper under review — author and affiliation information withheld
Left: overlaid contact forces on an L-shaped object and on a box. Right: eight distinct dexterous-hand grasps (top row) and eight distinct dual-humanoid hand-over configurations (bottom row) produced by Stein Projected ADMM.

Fig. 1: This paper proposes Stein Projected ADMM, a method that discovers diverse, feasible contact-rich manipulation strategies by combining Stein variational inference and Alternating Direction Method of Multipliers (ADMM). Illustrated are diverse solutions for single-object grasps (top) and dual-humanoid stable hand-over (bottom).

Abstract

Contact-implicit trajectory optimization formulates contact-rich manipulation as a single constrained program; however, that single program run collapses onto one local optimum out of many equally valid contact modes, grasps, or push directions. As a consequence, the resulting manipulation strategy is reluctant to change and sensitive to initialization. In order to promote robust manipulation, this paper investigates how contact-implicit solvers can discover diverse contact-rich strategies. Our approach derives a variation of Consensus Alternating Direction Method of Multipliers (ADMM) combined with Stein variational inference methods to output a set of distinct contact-rich solutions. We find that applying the Stein repulsive force to ADMM's split variable (rather than its primal form) allows for effective coverage over the set of feasible contact strategies without prematurely stalling the solver. We demonstrate the effectiveness of our approach on a variety of contact-rich manipulation tasks, including pushing, grasping, and multi-robot handover. Last, we find the proposed solver is simpler in form and capable of discovering unique contact modes when compared with existing solvers.

Methods

Seven panels showing the Stein Projected ADMM update on a complementarity problem: initial particles, reference repelled z, projected z at iterations 0 and 1, and the feasible, evenly spread particle set after 100 iterations.

Fig. 2: Overview of projected variable-splitting construction on a motivating complementarity problem with constraints 0 ≤ λ ⊥ φ ≥ 0, and objective f = ½∥[λ, φ]⊤ − [λ⋆, φ⋆]⊤∥22. The initial particles x are first split into duplicate z variables. The Stein repulsion is measured according to z to produce z̃ as a target reference which is projected onto the feasible set C. The x-update then tracks the objective while being pulled toward the feasible set C through the split variable.

Contact-implicit optimization encodes contact through complementarity constraints, so the feasible set is a union of lower-dimensional manifolds — one per contact mode. A single solve lands in one of them and discards every other equally valid strategy. Prior constrained-SVGD methods fold feasibility and diversity into the same update, so the repulsion term ends up competing with constraint satisfaction.

Our construction separates the two. The Stein repulsive force acts only on ADMM's split variable z, which is then exactly projected onto the feasible set at every outer iteration. The primal x-update therefore never sees an infeasible perturbation; it simply chases a target that is already feasible and already diverse. This requires no kernel-space quadratic program and no merit-function line search.

Results

Solver in Motion

Hardware Experiments

A strip of eight photographs of a dexterous robot hand grasping a green non-convex object with visibly different finger placements in each.

Diverse grasps on Allegro hand hardware solved with Stein-projected ADMM. We evaluate four-fingered quasi-static grasps of a non-convex object that runs in real-time (we refer the reader to the multimedia material). The executed grasp is chosen via an optimistic (best cost) heuristic from the solved candidate grasps.

Executing candidate grasps

One solve returns an ensemble of candidate grasps, and our method executes the best-quality grasp from that particle set. Below, four of the candidates are run on the hardware one at a time: each reaches the same quasi-static equilibrium on the same non-convex object, but commits to a different set of finger placements and contact forces. Each grasp is then disturbed by hand to check that it actually holds.