allenai/molmospaces

An end-to-end open ecosystem for robot learning

What it solves

MolmoSpaces provides a large-scale open ecosystem for robot manipulation and navigation, addressing the need for diverse, high-quality simulation environments and datasets to train and evaluate robotic policies. It enables the creation of massive amounts of synthetic data to help robots achieve zero-shot manipulation capabilities in the real world.

How it works

The project provides a framework for scene conversion, grasp generation, and benchmark evaluation across multiple simulators including MuJoCo, Isaac, and ManiSkill. It uses an asset manager to handle a vast library of objects (such as those from Objaverse) and scenes (including hand-crafted and procedurally generated environments). Users can define experiments via configuration files to generate data or evaluate policies on specific tasks like picking, opening, or closing objects.

Who it’s for

It is designed for robotics researchers and developers working on robot learning, manipulation, and navigation who need scalable simulation assets and standardized benchmarks for testing AI policies.

Highlights

  • Multi-Simulator Support: Assets are compatible with MuJoCo, Isaac, and ManiSkill.
  • Massive Asset Library: Includes over 129k converted Objaverse assets and thousands of procedurally generated scenes.
  • Integrated Benchmarking: Features a dedicated benchmark for atomic tasks (e.g., pick, open, close) with associated leaderboards.
  • Teleoperation Support: Includes integration with the TeleDex iPhone app for real-time robot control in simulation.
  • GPU Acceleration: Supports cuRobo for accelerated motion planning.

Related

  • Project
    ARISE-Initiative/robosuite

    A MuJoCo-powered simulation framework and benchmark suite for robot learning, providing modular tools to create environments and test AI control algorithms.

  • Project
    RoboVerseOrg/RoboVerse

    A unified platform, dataset, and benchmark designed to enable scalable and generalizable robot learning by integrating multiple simulation frameworks and datasets.

  • Project
    google-deepmind/mujoco

    A high-performance physics engine for the fast and accurate simulation of articulated structures, widely used in robotics and machine learning research.

  • Project
    RoboDojo-Benchmark/RoboDojo

    A unified sim-and-real benchmark for evaluating generalist robot manipulation policies across 60 tasks and five capability dimensions.

  • Project
    robfiras/loco-mujoco

    An imitation learning benchmark for whole-body control featuring diverse humanoid and quadruped environments with extensive motion capture datasets.