Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

20 Commits
 
 
 
 
 
 

Repository files navigation

Awesome Robotics Benchmarks

A curated collection of robotics benchmarks organized by domain with concise, comparable tables.

Counts (tasks / metrics / robot configs) are recorded as numbers; modalities are listed in detail; SIM shows the simulator backend (e.g., CoppeliaSim, SAPIEN, MuJoCo).

For the Simulation domain, the summary table compares capabilities: physics engine, renderer, GPU support, OS, and license.

human_demonstration — trajectories recorded from a human controlling the robot/simulator (e.g., teleop, VR).

motion_planner — trajectories generated/executed by a planner/IK/trajectory optimizer (e.g., MoveIt, RRT, CHOMP).

synthetic — trajectories or episodes created programmatically or by generative scripts/LLMs without a physics-aware planner or human.

NOTE: Currently this repo is underdevelopment and any pr is welcomed.

Domains


Manipulation

  • RLBench (2019): A large-scale learning environment featuring 100 unique vision-guided manipulation tasks of varying difficulty. RLBench provides multimodal observations (RGB, depth, segmentation, proprioception) from multiple camera angles and an infinite supply of demonstration trajectories generated via built-in motion planners.

    Resources: Paper | Website | Code

  • Meta-World (2019): Open-source benchmark of diverse single-arm tabletop manipulation tasks for multi-task and meta-RL.

    Resources: Paper | Website | Code

  • RoboSuite (2020): Modular MuJoCo-based simulator and benchmark suite for robot manipulation with standardized tasks and multi-robot support.

    Resources: Paper | Website | Code

  • CALVIN (2021): An open-source simulated benchmark for learning long-horizon language-conditioned robot manipulation tasks. It features a Franka Emika Panda arm in four tabletop environments and 34 distinct manipulation tasks with unconstrained language instructions

    Resources: Paper | Website | Code | Data

  • FurnitureBench (2023): Reproducible real-world furniture assembly benchmark with standardized hardware/setup, 3D-printed parts, large teleoperation dataset, and a matching simulator (FurnitureSim).

    Resources: Paper | Website | Code | Data

  • LIBERO (2023): A benchmark for lifelong robot learning featuring multitask manipulation with language conditioning. Focuses on knowledge transfer and generalization across diverse manipulation tasks.

    Resources: Paper | Website | Code | Data

  • RoboCasa (2024): A large-scale simulation framework for generalist household robots, focusing on diverse kitchen environments.

    Resources: Paper | Website | Code

  • RoboTwin (2024): A dual-arm manipulation benchmark and data-generation framework that uses generative 3D models and LLMs to create diverse bimanual task scenarios. RoboTwin provides a real-to-sim “digital twin” pipeline to generate varied object models and expert demonstrations, and an evaluation platform aligned with a real dual-arm robot (COBOT Magic platform). It combines simulated expert data with real-world teleoperated demos for coordinated two-arm tasks.

    Resources: Paper | Website | Code | Data

  • VLABench (2024): Language-conditioned manipulation benchmark with 100 categories emphasizing long-horizon reasoning and world knowledge.

    Resources: Paper | Website | Code

  • RoboCerebra (2025): A benchmark for evaluating high-level reasoning in long-horizon manipulation. RoboCerebra provides a large-scale simulation dataset of complex household tasks with extended subtask sequences, generated by GPT-based instruction decomposition and executed by human operators in simulatio. It emphasizes “System 2” planning skills (deliberative, goal-directed thinking) in vision-language-conditioned manipulation.

    Resources: Paper | Website | Data

Summary Comparison

Benchmark Subtype Task Count Metric Count Robot Configs Modality SIM Data Source Data Size
RLBench table-top 100 1 (task success rate) 1 RGB, depth, segmentation, proprioceptive CoppeliaSim (V-REP) motion_planner 100 tasks × infinite demos (motion planner can generate unlimited trajectories)
Meta-World table-top 50 1 1 proprioceptive MuJoCo unknown N/A
RoboSuite table-top, mobile-manipulation 9 2 10 RGB, depth, proprioceptive MuJoCo human_demonstration N/A
CALVIN table-top 34 1 (primary metric: success rate) 1 RGB, depth, proprioceptive, language, tactile PyBullet human_demonstration Hours of teleoperated play data (20K language instructions); 4 environments (A–D) with ~23K trajectories for training
FurnitureBench table-top 8 2 1 RGB, proprioceptive NVIDIA Isaac Gym / PhysX (Factory) via FurnitureSim human_demonstration ≈219.6 hours, 5100 successful demonstrations
LIBERO table-top 130 4 3 RGB, depth, proprioceptive, language RoboSuite (MuJoCo) human_demonstration 130+ tasks, 2500+ demonstrations
RoboCasa mobile-manipulation (household) 100 (25 atomic + 75 composite) 1 (success rate for task completion) 3 (supports single-arm mobiles, humanoids, quadruped-with-arm) RGB, proprioceptive PhysX (NVIDIA Omniverse) both 100+K demonstration trajectories (e.g. 50 human demos ×25 skills + 72K generated)
RoboTwin dual-arm 50 1 (task success rate) 1 RGB, depth, language sapien both Synthetic dataset (hundreds of expert demos) + limited real demos per task
VLABench table-top 100 2 1 RGB, language MuJoCo (dm_control) unknown not specified
RoboCerebra table-top 1,000 training tasks + 60 held-out tasks (1,060 total) 2 (sequence success rate and subtask completion) 1 RGB, language, proprioceptive RoboSuite (MuJoCo) both 100k+ trajectories (e.g. 50 human demos ×25 skills + synthetic expansions)

Locomotion

  • Gymnasium MuJoCo Locomotion (classic subset) (2016): Canonical MuJoCo continuous-control locomotion tasks widely used as RL baselines (e.g., Ant, HalfCheetah, Hopper, Walker2d, Humanoid).

    Resources: Paper | Website | Code

  • Brax (2021): JAX-based differentiable physics engine reproducing classic MuJoCo-style tasks for massively parallel training.

    Resources: Paper | Website | Code

  • HumanoidBench (2024): Simulated humanoid benchmark with dexterous hands spanning whole-body manipulation and locomotion; designed for high-DoF control research.

    Resources: Paper | Website | Code

Summary Comparison

Benchmark Subtype Task Count Metric Count Robot Configs Modality SIM Data Source Data Size
Gymnasium MuJoCo Locomotion (classic subset) - 11 1 6 RGB, depth, proprioceptive MuJoCo unknown N/A
Brax - 5 1 5 proprioceptive Brax (JAX) unknown N/A
HumanoidBench humanoid, whole-body 27 1 4 proprioceptive, RGB, tactile MuJoCo unknown N/A

Navigation

  • KITTI (2012): Large-scale real-world autonomous driving dataset and benchmark suite for perception and odometry.

    Resources: Paper | Website | Data

  • AI2-THOR (2017): Open-source Unity-based embodied AI platform with interactive indoor scenes; supports navigation (iTHOR/RoboTHOR) and manipulation (ManipulaTHOR).

    Resources: Paper | Website | Code | Data

  • Habitat Navigation (PointNav/ObjectNav/ImageNav) (2019): Photoreal embodied navigation benchmarks (PointNav, ObjectNav, ImageNav) run in Habitat-Sim with standard metrics like Success and SPL.

    Resources: Paper | Website | Code | Data

  • ALFRED (2020): Household instruction-following benchmark that maps egocentric vision + language to action sequences in AI2-THOR.

    Resources: Paper | Website | Code | Data

  • SocialGym 2.0 (2023): 2D ROS-based multi-agent social navigation simulator/benchmark with Python API and MARL baselines.

    Resources: Paper | Website | Code

Summary Comparison

Benchmark Subtype Task Count Metric Count Robot Configs Modality SIM Data Source Data Size
KITTI - 8 8 1 RGB, lidar, GPS_IMU - human_demonstration Varies by task; multi-sensor vehicle platform
AI2-THOR navigation+manipulation N/A N/A 2 RGB, depth, semantic Unity 3D synthetic iTHOR: 120 rooms; 2000+ objects; RoboTHOR: sim-real apartments
Habitat Navigation (PointNav/ObjectNav/ImageNav) - 3 2 1 RGB, depth, semantic Habitat-Sim synthetic large-scale episodes from HM3D/MP3D/Gibson splits (varies by year)
ALFRED navigation+manipulation 7 2 1 RGB, depth, semantic, language AI2-THOR 2.0 both 25,743 directives; 8,055 expert demos
SocialGym 2.0 social navigation 4 4 1 lidar, proprioceptive ROS-based 2D (PettingZoo/SB3 integration) unknwon N/A (simulated training/eval)

HRI

  • Assistive Gym (2019): Physics-based assistive robotics framework with human models and ADL tasks for safe robot assistance.

    Resources: Paper | Code

Summary Comparison

Benchmark Subtype Task Count Metric Count Robot Configs Modality SIM Data Source Data Size
Assistive Gym - 6 1 4 proprioceptive, RGB PyBullet N/A N/A

Safety

  • Safety Gym (2019): Constrained-RL benchmark with hazard/constraint costs across goal, push, and button tasks and multiple robots.

    Resources: Paper | Website | Code

  • safe-control-gym (2021): Unified safe learning-based control suite with constraints/disturbances on cart-pole and quadrotor systems.

    Resources: Paper | Website | Code

Summary Comparison

Benchmark Subtype Task Count Metric Count Robot Configs Modality SIM Data Source Data Size
Safety Gym - 18 3 3 lidar, proprioceptive MuJoCo unknown N/A (simulated training/eval)
safe-control-gym - 2 3 3 proprioceptive PyBullet unknown N/A (simulated training/eval)

Simulation

  • MuJoCo (2012): A fast, accurate physics engine for rigid-body simulation in robotics and control.

    Resources: Paper | Website | Code

  • SAPIEN (2020): Physics-rich robotics simulator with Vulkan renderer and PhysX 5, optimized for large-scale data generation.

    Resources: Paper | Website | Code

  • Genesis (2024): GPU-accelerated universal simulator with photorealistic rendering; targets high-throughput robot learning.

    Resources: Website | Code

  • Gazebo: Modular, open-source robotics simulator with pluggable physics and rendering backends.

    Resources: Website | Code

  • Isaac Lab: Open-source robot learning framework built on NVIDIA Isaac Sim (PhysX + RTX).

    Resources: Website | Code

  • PyBullet: Pythonic interface to Bullet physics; lightweight, widely used for robotics RL research.

    Resources: Website | Code

Summary Comparison

Simulation Physics Engine Renderer GPU Support OS License
MuJoCo MuJoCo OpenGL Linux, Windows, macOS Apache-2.0
SAPIEN NVIDIA PhysX 5 (CPU/GPU) Vulkan-based (SapienRenderer), ray tracing Linux MIT
Genesis Genesis (rigid/soft fluids) Photorealistic rasterization + ray tracing Linux, Windows, macOS Apache-2.0
Gazebo Plugins: DART (ref), Bullet, ODE, TPE; Chrono via plugin gz-rendering (OGRE/OGRE2; OptiX experimental) Linux, Windows, macOS Apache-2.0
Isaac Lab NVIDIA PhysX 5 (via Isaac Sim) NVIDIA Omniverse RTX / path tracing Linux, Windows BSD-3-Clause (Isaac Lab); Isaac Sim: Apache-2.0 repo + NVIDIA Omniverse EULA
PyBullet Bullet (rigid/soft body) OpenGL visualizer; TinyRenderer (CPU) Linux, Windows, macOS zlib

Generalist

  • ManiSkill (2021): SAPIEN-based benchmark targeting generalizable skills with four articulated-object and mobile manipulation tasks and a large LfD dataset.

    Resources: Paper | Website | Code | Data

  • EmbodiedBench (2025): A comprehensive benchmark to evaluate vision-driven embodied agents (multi-modal LLM-based) across both high-level and low-level task. EmbodiedBench spans four simulated environments – EB-ALFRED and EB-Habitat (high-level household tasks), and EB-Navigation and EB-Manipulation (low-level navigation and robotic manipulation) – comprising 1,128 diverse test instances in total. It also defines six capability-oriented evaluation subsets to assess commonsense reasoning, complex instruction understanding, spatial awareness, visual perception, long-horizon planning, etc.

    Resources: Paper | Website | Code | Data

Summary Comparison

Benchmark Subtype Task Count Metric Count Robot Configs Modality SIM Data Source Data Size
ManiSkill - 50 1 39 RGB, depth, pointcloud, proprioceptive SAPIEN (PhysX) motion_planner ~36,000 successful trajectories (~1.5M frames)
EmbodiedBench - 1128 (testing tasks across 4 envs) 2 (e.g., success rate and subgoal success) 4 RGB, language Multiple – AI2-THOR (Unity) for EB-ALFRED, Habitat-Sim for EB-Habitat/Navigation, and a robotics simulator for EB-Manipulation both (uses existing human-collected and synthetic tasks) 1,128 evaluation scenarios (drawn from ALFRED, Habitat, etc.)

Other

  • CARLA (2017): Open-source urban driving simulator and benchmark; standard CoRL’17 suite of goal-directed navigation tasks.

    Resources: Paper | Website | Code

  • D4RL (2020): Offline RL benchmark with standardized environments and fixed datasets across multiple domains.

    Resources: Paper | Website | Code | Data

Summary Comparison

Benchmark Subtype Task Count Metric Count Robot Configs Modality SIM Data Source Data Size
CARLA automous-driving N/A N/A N/A RGB, depth, semantic, lidar, GPS_IMU Unreal Engine (CARLA) unknown N/A (simulated scenarios)
D4RL Offline-RL 50 1 6 proprioceptive, RGB, lidar Multiple (MuJoCo, PyBullet, MiniGrid) both Multiple task-specific datasets

About

A curated, domain-organized list of robotics benchmarks—manipulation, locomotion, navigation, HRI, safety, simulation—summarized in concise, comparable tables.

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages