This work proposes Transition Occupancy Matching as a unifying principle to resolve policy and dynamics shifts within a single mathematical framework and introduces Occupancy-Matching Policy Optimization (OMPO), a novel algorithm that optimizes a surrogate objective explicitly correcting for transition discrepancies.
Abstract
Embodied agents must continuously adapt to the physical world using interaction data collected across varying timescales, controllers, and environmental conditions. However, standard reinforcement learning assumes stationary dynamics and on-policy data, a premise often violated in reality where physical parameters drift and historical data becomes heterogeneous. The central challenge lies in the compound distribution shift: replayed transitions follow an occupancy distribution that diverges fundamentally from the current physical reality, leading to biased value estimation and catastrophic learning collapse. In this work, we propose Transition Occupancy Matching as a unifying principle to resolve policy and dynamics shifts within a single mathematical framework. We introduce Occupancy-Matching Policy Optimization (OMPO), a novel algorithm that optimizes a surrogate objective explicitly correcting for transition discrepancies. By leveraging a dual reformulation with a sign-free logarithmic link, OMPO transforms the intractable matching problem into a stable min-max optimization, amenable to arbitrary reward structures. Crucially, OMPO integrates a distributional critic and a multimodal encoder with a small-scale local buffer, allowing the agent to anchor massive historical data to the immediate physical context for rapid adaptation. Extensive evaluations across diverse benchmarks—including MuJoCo locomotion, DM-Control, Meta-World, and high-fidelity Panda robot manipulation—demonstrate that OMPO consistently outperforms specialized baselines in stationary, domain-shifting, and non-stationary settings. By unifying distribution correction across policy and dynamics shifts, OMPO addresses a fundamental bottleneck in transfer learning, providing a robust algorithmic framework for continual adaptation in changing physical conditions.
Offline reinforcement learning (RL) must reconcile two competing requirements: policy updates should stay near dataset-supported actions to keep value estimates reliable, yet meaningful gains often require moving beyond the behavior distribution. We develop a geometric view of offline actor updates by modeling policies as a probability manifold endowed with a chosen metric geometry. Under this lens, a broad class of offline actor objectives can be interpreted as a single proximal policy improvement step (SPI), i.e., an implicit discretization of a manifold gradient flow induced by a critic-defined energy. Building on this insight, we propose multi-step proximal policy improvement (MPI), a plug-in refinement mechanism that composes sequential re-centered proximal steps. MPI enables controlled policy improvement beyond dataset support while retaining proximal control at each refinement. The framework accommodates multiple policy geometries and admits practical instantiations for deterministic and diagonal-Gaussian policies. Experiments on D4RL benchmarks show that small numbers of MPI refinements improve strong offline baselines, including TD3+BC, ReBRAC, and IQL, on many tasks. Focused diagnostics further distinguish re-centered refinement from fixed-objective update scheduling and characterize limitations under critic error.
A model-agnostic hybrid dynamics framework that blends a provably contracting nominal model with a flexible excursion model through an uncertainty-guided switching law is proposed, ensuring that each model operates within its reliability regime.
We develop the Continuous Distributed Coupled Policy Gradient (CDCPG) algorithm for cooperative reinforcement learning in networked Markov decision processes with continuous state and action spaces. Each agent maintains a local actor over a bounded graph neighborhood, and a localized least-squares temporal-difference critic evaluates a truncated action-value function through a spectral random-feature representation of the local transition kernel. The analysis makes four contributions. First, the truncated action-value function is constructed as a conditional expectation over the neighborhood, yielding a well-posed localized Bellman theory that removes the continuation-kernel mismatch of naive truncation arguments. Second, we expose a dimensional obstruction to temporal-difference stability for normalized random features and prove an unconditional excitation bound that reduces stability to a symmetric persistence-of-excitation condition, monitorable through an online matrix-concentration certificate. Third, under exponential spatial decay of agent interactions, the excitation condition, and smoothness of the objective, CDCPG drives an averaged per-agent stationarity measure to within any excess $\epsilon$ of an explicitly characterized approximation floor using $\widetilde{\mathcal{O}}(\epsilon^{-2})$ shared-oracle samples, and the excess dependence matches the smooth nonconvex first-order rate; per-agent computation and communication are governed by the neighborhood size rather than the network size. Fourth, an adaptive-locality rule selects the radius that balances truncation and graph-decay residuals against the target accuracy. Experiments on a networked linear-quadratic benchmark corroborate the locality and feature-dimension predictions.
This work proposes CoDrift, a compositional framework for one-step generative policy learning that combines three objective-level fields into a unified policy field that compares favorably with state-of-the-art methods and achieves the best average rank in both settings.
Effective model-based reinforcement learning in stochastic environments requires planning that accounts for predictive uncertainty. Propagating full state distributions analytically offers a principled way to do this, but has traditionally required restrictive policy or reward structures to remain tractable. Consequently, modern deep reinforcement learning has largely retreated to either stochastic sampling, which introduces significant target variance, or deterministic point estimates that ignore predictive covariance entirely. We investigate whether distribution-aware planning is possible without these constraints. Using a quadratic action-value parameterization, we first reduce the Bellman backup to an expectation over the state-value function alone; the key idea is then a compatibility principle between the predictive transition distribution and the value function class, under which this expectation is analytic in the distribution's moments. We instantiate this principle with a Gaussian transition model paired with a radial-basis value function, yielding a closed-form backup that propagates both predictive mean and covariance. Empirically, our approach reduces target variance and yields well-calibrated predictive uncertainty under stochastic observations in continuous control, providing a principled framework for planning with learned distribution models.
The Safe Deep Successor Representation is proposed, a novel method that allows quick retraining of policies towards new cost structures and is competitive on a simple navigation task while being considerably more flexible.
Michaela Girstl, Alexander Mattick, Christopher Mutschler· Trans. Mach. Learn. Res.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
