Distributionally Robust Control via Stein Variational Inference for Contact-rich Manipulation

Venue: RSS 2026 (Sydney, Jul 13–17) · Session: Manipulation 2 · paper #61 Authors: Hrishikesh Sathyanarayan, Victor Vantilborgh, Harish Ravichandar, Tom Lefebvre, Ian Abraham arXiv: 2605.19029 · program page

Summary compiled from the arXiv paper (v2); all numbers quoted from the paper. Trend context: RSS 2026 survey.

Figure 1: within-hand dynamic positioning of a cup with unknown mass distribution and friction — the robot slides the cup along a tray from initial (red) to goal (green) pose via controlled sliding; inset shows the Stein DRO planner's inferred trajectory particles under parameter uncertainty.

Problem

Contact-rich manipulation must cope with latent physical parameters (mass distribution, inertia, friction) that are hard to observe: data-driven policies need large-scale training and degrade with few samples, while classical worst-case distributionally robust (DRO) controllers become overly conservative because contact dynamics are discontinuous and mode-dependent, sacrificing task performance.

Method

SV-DRO (Stein Variational Distributionally Robust Optimizer) casts manipulation MPC as a distributionally robust optimization under a KL ambiguity set, but instead of guarding against a fixed worst case, it uses Stein variational gradient descent to evolve deterministic parameter particles from a prior toward a task-aware posterior induced by the optimality gap — concentrating on parameter realizations most consequential for task success. The controller runs receding-horizon rollouts through a surrogate spring-like contact model, executes H controls, then applies SVGD updates at the (low) feedback frequency, so no persistent state feedback or online system identification is required. Kernel choice (RBF, k=1, IMQ) is studied; the whole loop runs with constant computational scaling on an RTX 3080.

Results

Over 32 simulated trials with bounded-uniform parameter priors (Table I), SV-DRO reaches 93.25% / 84.38% success (≤10 cm / ≤1 cm) on bimanual Push-T and 100% / 100% on within-hand dynamic positioning, vs. DuST-MPC (71.88%/59.38% and 84.38%/46.88%), EMPPI (43.75%/31.25% and 59.38%/40.63%), naive MPC (28.13%/18.75% and 31.25%/25%), and conventional DRO (3.13%/0% and 34.38%/28.13%) — the claimed up-to-3x robustness gain. On hardware, a continual 3.5-minute within-hand run with random cup perturbations achieves the 1.5 cm tolerance in approximately 100% of trials; all kernels give 100% success on the kernel study with IMQ fastest (8.78 ± 1.34 s vs. 12.93 s RBF).

Significance

A model-based counterpoint to the session's learning-heavy entries: task-sensitive uncertainty shaping recovers reliability that worst-case DRO throws away, with no training data at all. Related threads: Review-Dexterous-Manipulation · Review-VLA-Evaluation.

← Back to RSS 2026 survey · RSS-2026-Papers · Home