allenai/molmospaces

What it solves

MolmoSpaces provides a large-scale, open ecosystem for training and evaluating robot manipulation and navigation. It addresses the gap in available high-quality, diverse simulation assets and standardized benchmarks for robotic AI, enabling researchers to generate synthetic data and test policies across multiple simulators.

How it works

The project provides a comprehensive library of assets—including thousands of 3D objects (from Objaverse and hand-crafted sets) and diverse scenes (procedurally generated or LLM-generated)—that are compatible with MuJoCo, Isaac Sim, and ManiSkill. It includes a data generation pipeline that uses experiment configurations to define robots, tasks, and scenes, as well as a benchmark evaluation system to measure performance on atomic robotic tasks.

Who it’s for

Robotics researchers and AI developers working on embodied intelligence, specifically those focusing on robot manipulation, navigation, and the creation of synthetic training data for robotic policies.

Highlights

• Multi-Simulator Support: Assets are usable across MuJoCo, Isaac, and ManiSkill.
• Large-Scale Asset Library: Includes over 129k converted Objaverse assets and thousands of hand-crafted indoor objects.
• Diverse Scene Generation: Supports procedurally generated scenes and LLM-generated environments.
• Integrated Benchmarking: Provides a dedicated benchmark for evaluating atomic tasks with a leaderboard.
• Teleoperation Support: Integration with the TeleDex app for phone-based robot control.