Research map / Reinforcement learning / Policy gradient and actor-critic

SRL-MPC: Shape-Aware Reinforcement Learned Model Predictive Control

NeurIPS 2026 · Reinforcement learning › Policy gradient and actor-critic · approach: Model-based RL with guarantees · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SRL-MPC, which formulates shape-aware degree-2 HOCBF constraints from support-function-based geometric separation features and uses PPO to tune MPC weights and safety distance online for navigating crowds of polygonal robots.

Problem

Safe, real-time navigation of many robots with arbitrary convex/nonconvex shapes in dense crowds, where prior methods assume circular shapes, homogeneous robots, or hand-tuned parameters.

Approach

Encodes pairwise shape-aware separation as geometric separation features (GSFs) via support function transformation, imposes degree-2 HOCBF constraints as a soft residual penalty in a decomposed local MPC, and uses an RL policy that reads GSFs and outputs bounded increments to MPC weights (w_p, w_u) and safety distance rather than end-to-end control.

Key results

• At 25 robots SRL-MPC achieves 92.0% success vs. 21.0% for the best external baseline (SARL) and 7.0% for ORCA.
• At 10 robots SRL-MPC reaches 100.0% success with 0 collisions, versus 95.0% for AVOCADO.
• Ablation at 25 robots: SRL-MPC 92.0% success vs. rule_mpc 75.0%, manual_mpc 68.0%, dist_mpc 7.0%.

Surprising findings

• The static distance-margin penalty (dist_mpc) collides in 84-92% of episodes at 20-25 robots despite the fastest arrival times.
• Under 0.05 m neighbor-position noise success drops to 83% and at 0.2 m to 1%, showing high sensitivity to perception error.

Limitations

• Higher per-robot compute than reactive baselines (about 10 ms at 25 robots), and the paper states real-time feasibility only at the 10 ms level.
• Evaluated only in randomized convex polygon simulation; nonconvex, heterogeneous dynamics and full perception pipelines are future work.
• Compared against OBCA is missing; sensitive to neighbor position noise and action delay.

Experimental setup

Datasets: IR-SIM randomized random_polygon scenarios (simulated)

Benchmarks: IR-SIM random_polygon scenario (N=10,15,20,25)

Metrics: success rate collision rate timeout rate navigation time speed path length

Baselines: ORCA AVOCADO VO-polytope SARL RL-RVO dist_mpc manual_mpc rule_mpc

Builds on: control barrier function PPO GEOS differential-drive MPC path tracking

Closest papers by meaning

• TD-MPC2: Scalable, Robust World Models for Continuous Control (ICLR 2024 · similarity 0.84)
• Position: Automatic Environment Shaping is the Next Frontier in RL (ICML 2024 · similarity 0.83)
• HASARD: A Benchmark for Vision-Based Safe Reinforcement Learning in Embodied Agents (ICLR 2025 · similarity 0.83)
• Deep Reinforcement Learning for Robotics: A Survey of Real-World Successes (AAAI 2025 · similarity 0.83)
• METRA: Scalable Unsupervised RL with Metric-Aware Abstraction (ICLR 2024 · similarity 0.83)
• CarPlanner: Consistent Auto-regressive Trajectory Planning for Large-Scale Reinforcement Learning in Autonomous Driving (CVPR 2025 · similarity 0.83)