Survey on Physics Engines, Simulation Frameworks, and Benchmarks for Robot Learning

This is the accompanying repository from the paper "Survey on Physics Engines, Simulation Frameworks, and Benchmarks for Robot Learning." This repository tracks updates in the field following our definition of Physics Engines, Simulation Frameworks, and Benchmarks in the robot learning domain. We include a curated list of tools for each type of software, but more exhaustive lists can be found in other repositories (see the best-of-robot-simulators for example).

Comparison of features across physics engines typically used in Robot Learning

This table provides a side-by-side overview of the core capabilities and trade-offs offered by the most popular physics engines in robot learning. Each column highlights a key dimension—accuracy, GPU acceleration, joint configurability, soft-body support, contact solver behavior, documentation quality, open-source status, sensor modalities, multi-agent scalability, and ease of physics parameter tuning for domain randomization. Use it to quickly spot which engines are best suited for high-precision manipulation tasks (e.g., MuJoCo, Chrono), GPU-parallelized large-scale training (e.g., PhysX, MJX), or rich soft-body interaction (e.g., Bullet, Chrono).

Physics Engine	Accuracy	GPU	6-DoF Configurable Joint	Soft Body Support	Contact Solver Characteristics	Docs	Open Source	Sensor Support	Multi-agent Scalability	Physics Params Support
Mujoco	High	❌	❌	❌	No internal forces, Robust, Convergence guarantees	Comprehensive	✅	Limited (Camera, no LiDAR)	Limited	High
MJX	Medium	✅	❌	❌	No internal forces, Robust, Convergence guarantees	Comprehensive	✅	Limited (inherits MuJoCo's model)	Support with JAX	Medium
Bullet	Medium	~	✅	✅	Hard contacts	Well-documented	✅	Moderate	CPU bottlenecks	Medium
Havok	Low	❌	✅	✅	Not strict convergence	Limited documentation	❌	Very limited	Limited	Low
ODE	Low	❌	❌	❌	Hard contacts	Partially documented	✅	Very limited	Limited	Medium
PhysX	Medium	✅	✅	✅	Hard contacts	Well-documented	✅	Moderate	High	Medium
Dart	High	❌	❌	❌	Hard contacts	Well-documented	✅	Good	Good	High
Genesis	Medium	✅	❌	✅	~	Partially documented	✅	Limited	Moderate	Medium
Chrono	High	✅	✅	✅	Hard contacts	Well-documented	✅	Comprehensive	Moderate	High

Comparison of Simulation Frameworks for Robot Learning

Simulation frameworks wrap a physics engine in high-level tooling for robot modeling, environment construction, sensor emulation, and experiment management. This table compares their native interoperability with learning libraries, support for curriculum and domain randomization (both physical and visual), rendering fidelity, proven sim-to-real workflows, and out-of-the-box support for locomotion, manipulation, and navigation tasks. “✅” means full built-in support, “~” indicates partial or emerging capabilities, and “❌” means none. Choose a framework that matches your workflow needs—whether that’s top-tier rendering for vision-based manipulation (e.g., Unity, IsaacSim), flexible randomization APIs for robust policy training (e.g., MuJoCo Playground, SAPIEN), or strong multi-agent and navigation features (e.g., Habitat, CARLA).

Frameworks	IsaacSim	SAPIEN	Unity	CoppeliaSim	Gazebo	CARLA	Habitat	AI2Thor	Robosuite	iGibson	MuJoCo Playground	PyBullet
Interoperability with Learning Frameworks	✅	✅	✅	~	❌	✅	✅	✅	✅	✅	✅	~
Domain Randomization Curriculum	~	~	❌	❌	❌	❌	❌	❌	❌	❌	✅	❌
Visual Domain Randomization	✅	✅	❌	❌	❌	✅	✅	✅	~	❌	❌	~
High Fidelity Rendering Support	✅	✅	✅	❌	❌	✅	✅	~	✅	✅	❌	❌
Sim-to-Real Track Record	✅	✅	❌	~	~	❌	~	~	~	❌	❌	✅
Locomotion Support	✅	~	~	~	~	❌	❌	❌	❌	~	✅	✅
Manipulation Support	✅	✅	~	✅	~	❌	❌	❌	✅	✅	✅	✅
Navigation Support	~	~	✅	~	✅	✅	✅	✅	❌	✅	❌	~

Comparison of Benchmarks for Robot Learning

Benchmarks offer standardized tasks and evaluation protocols so that algorithms can be rigorously compared. In this table you’ll find each benchmark’s target domains (manipulation, locomotion, navigation), whether it offers diverse reward structures, environment complexity, built-in sim-to-real evaluations, visual robustness tests, and the overall breadth of tasks included. Use this to pick a benchmark aligned with your research focus—whether that’s long-horizon compositional manipulation (e.g., CALVIN, ManiSkill), multi-domain challenges (e.g., RoboHive, HumanoidBench), or navigation-specific suites (e.g., Habitat Challenge).

Benchmark	Domains Supported	Reward Diversity	Environment Complexity	Sim2Real	Visual Robustness	Diversity of Tasks
Mujoco Playground	Manipulation, Locomotion	~	❌	~	~	~
DeepMind Control	Locomotion	❌	❌	❌	❌	❌
OGBench	Manipulation	~	❌	❌	~	✅
RoboHive	Manipulation, Locomotion	✅	✅	✅	~	✅
Meta-World	Manipulation	✅	~	~	~	✅
RlBench	Manipulation	✅	✅	✅	✅	✅
ManiSkill	Manipulation	✅	✅	~	✅	✅
DMC-VB	Manipulation	~	❌	❌	~	~
DMC-GB2	Manipulation, Locomotion	~	❌	❌	~	~
VD4RL	Manipulation	~	~	❌	✅	~
RL-ViGen	Manipulation	~	~	❌	✅	✅
CALVIN	Manipulation	✅	✅	✅	✅	✅
Colosseum	Manipulation	~	~	~	~	~
HumanoidBench	Manipulation, Locomotion	❌	✅	❌	✅	✅
LocoMuJoCo	Locomotion	✅	❌	~	❌	✅
Habitat Challenge	Navigation	✅	✅	✅	✅	~

Task-Specific Resources for Robot Learning

Beyond general simulation comparisons, it's often helpful to look at task-specific research and implementations. This section curates representative works and public repositories categorized by robotic capabilities such as locomotion, manipulation, and navigation. Each table includes references to learning-based approaches, real-world deployments, and sim-to-real workflows for specific robot platforms.

Locomotion Resources for Legged Robots

	A1	Go1	H1	Anymal	spot	cassie	solo12
MuJoCo	A Walk in the Park: Learning to Walk in 20 Minutes With Model-Free Reinforcement Learning Reinforcement Learning with Evolutionary Trajectory Generator: A General Approach for Quadrupedal Locomotion	SLIM: Sim-to-Real Legged Instructive Manipulation via Long-Horizon Visuomotor Learning	unitree_rl_gym			Reinforcement Learning for Versatile, Dynamic, and Robust Bipedal Locomotion Control Sim-to-Real Learning for Bipedal Locomotion Under Unsensed Dynamic Loads Learning Locomotion Skills for Cassie: Iterative Design and Sim-to-Real
IsaacGym/IsaacSim	Unified Locomotion Transformer with Simultaneous Sim-to-Real Transfer for Quadrupeds CPG-RL: Learning Central Pattern Generators for Quadruped Locomotion HYBRID INTERNAL MODEL: LEARNING AGILE LEGGED LOCOMOTION WITH SIMULATED ROBOT RESPONSE Robot Parkour Learning Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning	Robot Parkour Learning	unitree_rl_gym Humanoid Parkour Learning A Unified and General Humanoid Whole-Body Controller for Fine-Grained Locomotion	Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning		Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning	CaT: Constraints as Terminations for Legged Locomotion Reinforcement Learning
PyBullet	Robust High-Speed Running for Quadruped Robots via Deep Reinforcement Learning LEARNING VISION-GUIDED QUADRUPEDAL LOCOMOTION END-TO-END WITH CROSS-MODAL TRANSFORMERS GenLoco: Generalized Locomotion Controllers for Quadrupedal Robots	GenLoco: Generalized Locomotion Controllers for Quadrupedal Robots		GenLoco: Generalized Locomotion Controllers for Quadrupedal Robots			Model-free reinforcement learning for robust locomotion using demonstrations from trajectory optimization
RaiSIM				Learning Agile and Dynamic Motor Skills for Legged Robots Learning Quadrupedal Locomotion over Challenging Terrain RLOC: Terrain-Aware Legged Locomotion using Reinforcement Learning and Optimal Control			Controlling the Solo12 Quadruped Robot with Deep Reinforcement Learning
Gazebo	Robust High-Speed Running for Quadruped Robots via Deep Reinforcement Learning	Robust Localization, Mapping, and Navigation for Quadruped Robots			Spot_simulation

Note: We only list the latest commercially available or open‐source versions of legged robots that multiple research labs have used. Contributions are welcome via pull request.

Learning-Based Manipulation Resources for Dexterous Robots

Simulator	Franka	Humanoid robot Digit	ANYmal + arm	ANYmal (using feet for manipulation)	Unitree Go2 + arm	Unitree Go1 + arm	Unitree B1 + arm	Xbot	UR5e Robot	Berkeley BLUE Robot arm	Shadow Dexterous Hand	Fourier GR-1 Humanoid Robot	A2Single	DEX-EE Hand	ALOHA	Unitree H1	Unitree G1	Franka Emika	Toyota HSR robot	Unitree Aliengo + Z1 arm	Aliengo (Using feet for manipulation)
MuJoCo	Curriculum Design and Sim2Real Transfer for Reinforcement Learning in Robotic Dual-Arm Assembly Learning to Manipulate Anywhere: A Visual Generalizable Framework For Reinforcement Learning	Sim-to-Real Learning for Humanoid Box Loco-Manipulation								DoorGym: A Scalable Door Opening Environment and Baseline Agent	Solving Rubik's Cube with a Robot Hand Learning Dexterous In-Hand Manipulation Learning to Solve a Rubik's Cube with a Dexterous Hand			DemoStart: Demonstration-led auto-curriculum applied to sim-to-real with multi-fingered robots	Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware		FALCON: Learning Force-Adaptive Humanoid Loco-Manipulation
IsaacGym/IsaacSim			Learning to Open and Traverse Doors with a Legged Manipulator Guided Reinforcement Learning for Robust Multi-Contact Loco-Manipulation	Pedipulate: Enabling Manipulation Skills using a Quadruped Robot's Leg	UMI on Legs: Making Manipulation Policies Mobile with Manipulation-Centric Whole-body Controllers	Deep Whole-Body Control: Learning a Unified Policy for Manipulation and Locomotion	Learning Force Control for Legged Manipulation Visual Whole-Body Control for Legged Loco-Manipulation	HYPERmotion: Learning Hybrid Behavior Planning for Autonomous Loco-manipulation				Sim-to-Real Reinforcement Learning for Vision-Based Dexterous Manipulation on Humanoids				Unleashing Humanoid Reaching Potential via Real-world-Ready Skill Space	AMO: Adaptive Motion Optimization for Hyper-Dexterous Humanoid Whole-Body Control				Learning Visual Quadrupedal Loco-Manipulation from Demonstrations
PyBullet										Transporter Networks: Rearranging the Visual World for Robotic Manipulation
RaiSIM
Gazebo
SAPIEN / ManiSkill													Learning Generalizable Manipulation Policy with Adapter-Based Parameter Fine-Tuning						Ag2Manip: Learning Novel Manipulation Skills with Agent-Agnostic Visual and Action Representations	Learning Generalizable Manipulation Policy with Adapter-Based Parameter Fine-Tuning	Learning Generalizable Manipulation Policy with Adapter-Based Parameter Fine-Tuning
Unity										DoorGym: A Scalable Door Opening Environment and Baseline Agent
CoppeliaSim
iGibson
Robosuite	Curriculum Design and Sim2Real Transfer for Reinforcement Learning in Robotic Dual-Arm Assembly

Note: We only list the latest commercially available or open-source versions of legged robots that have been used by multiple research labs. If you know of any additional papers or repositories that are not shown here, feel free to open a pull request.

Changelog

This section will track all major updates to the engines, frameworks, and benchmarks covered in this survey. Whenever a new major version is released, capabilities are added or changed, or entirely new entries appear, record the date, the tool affected, and a short description of what’s new.

Date	Tool	Change Description	Notes / Links
June 2025	NA	Initial Commit of Resources after paper submission	–

Entries will be added here as releases are announced or new tools are included.