cs.LGNov 1, 2026

Iterative Gumbel Planning for Continuous Control

Authors: Shaohuai Liu, Weirui Ye, Yilun Du, Le Xie

Organizations: Texas A&M University · CSAIL, MIT · Harvard University

Abstract

Search-based planning remains a powerful policy improvement paradigm in reinforcement learning, driving milestone successes in discrete domains. However, extending this paradigm to continuous control exposes a structural mismatch: search evaluates a finite set of action candidates, whereas continuous RL typically relies on Gaussian policies. We show that winner-action distillation into a continuous Gaussian induces a projection error and a stepwise variance-expanding gradient tendency when the selected action is sufficiently far from the current mean. Empirically, this tendency is associated with elevated terminal variance and a return gap on the evaluated tasks. To address this, we propose Iterative Gumbel Planning (IGP). By adopting a per-dimension factorized categorical policy, IGP naturally aligns the policy support with the search candidates, eliminating the projection error and reducing the residual approximation to a bounded discretization error. To bypass the combinatorial explosion of high-dimensional discrete action spaces, IGP pairs Gumbel Top-KK candidate sampling with an MPPI-style multi-round iterative search to refine the policy at each state. Empirically, IGP achieves state-of-the-art aggregate performance and sample efficiency on DMControl and HumanoidBench. Furthermore, it produces more stable action sequences with lower temporal variance, demonstrating its practical potential for real-world robotic deployment.

Figures & tables

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Sep 18, 2026

GEM-MPC: Balancing Exploration and Exploitation through Expert-Guided Planning

Effective exploration in high-dimensional continuous control remains a central challenge in reinforcement learning. Planning-based methods address this by combining online planning with learned policies and value functions, but their components can become misaligned during training: learned sampling policies may diverge from planner behavior, while planning distributions stored in replay become stale as the model and value function evolve. Reanalysis can refresh these targets, but at substantial computational cost. We propose GEM-MPC, an MPPI-based reinforcement learning method that improves the interaction between planning and learning. GEM-MPC uses MPPI to combine a policy trained to clone the planner with a KL-regularized policy that explores around it, providing complementary exploitation and guided exploration within planning. We further introduce Gated Prior Distillation, which selectively learns from stored planning distributions only when they provide a better target than the current prior, reducing the impact of stale planning data without requiring full reanalysis. Across continuous-control benchmarks, GEM-MPC consistently outperforms existing planning-based baselines under lower computational budgets.
Feb 21, 2026

In-Context Planning with Latent Temporal Abstractions

Planning-based reinforcement learning for continuous control is bottlenecked by two practical issues: planning at primitive time scales leads to prohibitive branching and long horizons, while real environments are frequently partially observable and exhibit regime shifts that invalidate stationary, fully observed dynamics assumptions. We introduce I-TAP (In-Context Latent Temporal-Abstraction Planner), an offline RL framework that unifies in-context adaptation with online planning in a learned discrete temporal-abstraction space. From offline trajectories, I-TAP learns an observation-conditioned residual-quantization VAE that compresses each observation-macro-action segment into a coarse-to-fine stack of discrete residual tokens, and a temporal Transformer that autoregressively predicts these token stacks from a short recent history. The resulting sequence model acts simultaneously as a context-conditioned prior over abstract actions and a latent dynamics model. At test time, I-TAP performs Monte Carlo Tree Search directly in token space, using short histories for implicit adaptation without gradient update, and decodes selected token stacks into executable actions. Across deterministic MuJoCo, stochastic MuJoCo with per-episode latent dynamics regimes, and high-dimensional Adroit manipulation, including partially observable variants, I-TAP consistently matches or outperforms strong model-free and model-based offline baselines, demonstrating efficient and robust in-context planning under stochastic dynamics and partial observability.
Oct 19, 2026

Long-Horizon Multi-Object Rearrangement via Distilled Tree Search over Continuous Actions

Despite recent advances in policy learning, long-horizon manipulation remains difficult because learned policies must avoid compounding errors while preserving future feasibility. While search-based planning can explicitly reason over such future consequences, it quickly becomes expensive in high-dimensional continuous action spaces. This raises a central question: how can long-horizon reasoning from search be transferred into policies in continuous action spaces? To this end, we propose DISCO: Distilled Tree Search over Continuous Actions. DISCO uses progressive-widening MCTS in physics simulation to generate long-horizon plans. During search, a learned policy proposes continuous actions, while a learned value function scores actions by their expected progress toward task completion. Using GPU-parallelized simulation, DISCO periodically distills MCTS-generated plans into the policy-value model, improving the action proposals and value estimates used in subsequent searches. Finally, the learned policy-value model can be directly deployed without the need for simulator-based search at test time. Across simulated and real-world rearrangement tasks, DISCO outperforms behavior cloning, reinforcement learning, and search-based baselines.