Multi-Agent Reinforcement Learning for Robotics (2026): Multi-Robot Learning, Safety & Simulators

What Is Multi-Agent Reinforcement Learning for Robotics?

Multi-agent reinforcement learning for robotics trains several robots, or several parts of one robot, to learn coordinated behaviour. It is used for decentralized navigation and collision avoidance, cooperative manipulation, teams of different robot types and locomotion with multiple bodies. The setting adds practical difficulties to ordinary MARL: partial sensing, safety constraints, real-time limits and the gap between simulation and hardware. This page collects surveys, navigation and safety papers, heterogeneous-agent and dexterous-manipulation work, and the simulators and code used to test them. It is a reading list, not a ranking, and most results come from simulation.

Surveys & Reviews

4 entries

Overviews of how multi-agent RL is applied to teams of robots.

01Distributed Reinforcement Learning for Robot Teams: A ReviewReview of model-free multi-agent RL for multi-robot cooperation, organized by independent learning, centralized critic, value decomposition and communication learning. 02Multi-Agent Deep Reinforcement Learning for Multi-Robot Applications: A Survey (Sensors)Survey of multi-agent deep RL for multi-robot applications, available on PubMed Central. 03Deep Reinforcement Learning for Multi-Agent Systems: A Review of Challenges, Solutions and ApplicationsReviews challenges such as non-stationarity and partial observability, with applications that include robotics. 04Multi-agent Reinforcement Learning: A Comprehensive SurveyGeneral MARL survey that gives the background for the robotics-specific work.

Decentralized Navigation & Collision Avoidance

3 entries

Learned policies that let robots move among each other without a central planner.

05Decentralized Non-communicating Multiagent Collision Avoidance with Deep RLCollision avoidance for agents that cannot observe each other's goals or communicate. 06Towards Optimally Decentralized Multi-Robot Collision Avoidance via Deep RLDecentralized collision avoidance policy for multiple robots that do not observe each other's states and intents. 07Fully Distributed Multi-Robot Collision Avoidance via Deep RLDecentralized policy that maps raw sensor measurements directly to steering commands.

Safety, Heterogeneous Teams & Manipulation

3 entries

Methods for constraints, teams of different robots, and coordinated joints or hands.

08Multi-Agent Constrained Policy OptimisationPaper on policy optimisation with safety constraints in multi-agent RL, where each agent must also respect the constraints of others. 09Heterogeneous-Agent Reinforcement Learning (HARL)Cooperative MARL methods that do not rely on parameter sharing, so agents can be different. 10Bi-DexHands: Towards Human-Level Bimanual Dexterous Manipulation with RLBimanual dexterous manipulation benchmark that treats the cooperating hands and joints as heterogeneous agents.

Simulators, Benchmarks & Code

8 entries

Environments for training and testing multi-robot policies, with latest releases as of 15 November 2026.

11VMAS: A Vectorized Multi-Agent Simulator for Collective Robot LearningPaper on a vectorized simulator for multi-robot coordination problems. 12VMAS (GitHub)Vectorized differentiable simulator for multi-agent RL benchmarking. Latest release VMAS-1.5.2 (10 November 2025). 13NVIDIA Isaac LabUnified framework for robot learning with multi-physics and renderer support. Latest release v3.0.0 (20 October 2026). 14MuJoCoGeneral purpose physics simulator widely used in robot learning. Latest release 4.0.1 (11 November 2026). 15Multi-Agent Quadruped EnvironmentMulti-agent quadruped environment that supports learning of locomotion control or high-level planning. 16DexterousHandsLibrary of dual dexterous hand manipulation tasks built on Isaac Gym. 17HARL (GitHub)Official PyTorch implementation of the HARL algorithms. 18PettingZooStandard API for multi-agent reinforcement learning environments.

5 entries

Related pages on this site.

19Multi-Agent Reinforcement LearningGuide to multi-agent reinforcement learning (MARL): foundations, core algorithms such as MADDPG, QMIX and MAPPO, communication, benchmarks like SMAC and Melting Pot, and libraries including PettingZoo, BenchMARL and JaxMARL. 20AI Agent-Powered Swarm RoboticsGuide to AI agent-powered swarm robotics for 2026: LLM-driven multi-robot collaboration research, drone swarm tools, ROS agent frameworks, MCP, simulators and swarm programming languages, with primary sources. 21Swarm IntelligenceDirectory of swarm intelligence algorithms and frameworks in 2026: particle swarm optimization, ant colony optimization, artificial bee colony and firefly. 22Robot Action ModelsRobot action models in 2026: an editorial Top 15 for learning new tasks, including Gemini Robotics 2, π0.7, Helix 2.5, GR00T N1.7 and Isaac 0.5, with sources. 23Autonomous Mobile RobotsAutonomous mobile robots (AMRs) in 2026: vendors, ROS 2 Nav2 and SLAM software, fleet interoperability with VDA 5050 and Open-RMF, and safety standards.

Frequently Asked Questions

What is multi-agent reinforcement learning for robotics?

It is reinforcement learning in which several robots, or several parts of one robot such as joints and hands, learn policies together. The aim is coordination, such as moving through a crowd, carrying a load together or controlling a multi-fingered hand.

Why use MARL instead of a central planner?

Exact multi-robot planning often does not scale with the number of robots, and a central planner needs reliable communication. Learned decentralized policies act on each robot's own observations, as in the collision avoidance papers listed here, but they must handle other robots that are also learning.

What are the main challenges?

Other learning robots make the environment non-stationary, each robot sees only part of the scene, and real robots must respect safety limits. The multi-agent constrained policy optimisation paper addresses safety, and HARL addresses teams where robots are not identical.

How do I move from simulation to real robots?

Most of the work listed here is done in simulation first, in tools such as VMAS, Isaac Lab and MuJoCo. Moving to hardware adds sensing noise, latency and safety requirements. Keep safety-critical control and emergency stops in conventional software and test incrementally.

Where should I start?

Read a survey such as the robot-team review to learn the main approaches, try a simulator such as VMAS or Isaac Lab, and then use a MARL library to reproduce a benchmark. The companion Multi-Agent Reinforcement Learning page on this site lists the core algorithms and libraries.