Posterior-Driven Heuristic Support Adaptation in a Probabilistic Treatment of Real2Sim2Real for Vision-Driven Deformable Linear Object Manipulation

Abstract

Likelihood-free inference (LFI) enables system identification in complex tasks via closed-box modelling, abstracting nonlinearity and stochasticity, and infers a domain distribution for adapting agents to parametric deployment conditions. LFI assumes an arbitrary support for sampling, which remains fixed as the initial generic prior is refined to increasingly descriptive posteriors. Misspecified support can therefore yield suboptimal yet overconfident posteriors. We address this issue by using the posterior of an inference step to guide the adaptation of the support using three illustrative heuristics: EDGE, MODE, and CENTRE. Each heuristic interprets the updated belief and enables support adaptation alongside posterior inference. For illustrative purposes, we first study misspecified support in LFI and evaluate the utility of our heuristics using stochastic dynamical benchmarks. We then evaluate posterior-driven heuristic support adaptation for parameter inference and policy learning in a dynamic deformable linear object (DLO) manipulation task. Inference results in a finer length and stiffness classification for a parametric set of DLOs. When the resulting posteriors are used as domain distributions for sim-based policy learning, they lead to more robust object-centric agent performance. Note to Practitioners—This paper was motivated by the difficulty of defining the parameter space, or domain support, for the probabilistic inference of the length and stiffness of a deformable linear object through vision. This definition is nontrivial when inferring the distribution that will calibrate a simulation for policy learning, as we may lack intrinsic knowledge of the real-world system and how real physical properties correspond to the simulation parameters. Current methods assume an arbitrary ‘best guess’ support, but this is susceptible to misspecification for systems of increased complexity and stochasticity. We study how the inferred posteriors may indicate the need to adapt the support to more accurately reflect the range of possible parameterisations. We define a set of heuristics to iteratively adapt the support in tandem with posterior inference. Our experiments show that this increases the reliability of both parameter inference and agent skill adaptation. Our heuristics can be used as baselines for future methods that explore more formal approaches to support adaptation, or automatically extract the rule that should guide it. The potential impact of our methodology spans from closing the ‘reality gap’ to robustly conditioning autonomous agents for an estimated context under unknown system dynamics.