Summary

MolmoSpaces is a comprehensive, open-source ecosystem for robot manipulation and navigation research. It addresses the need for large-scale, reusable simulation assets and standardized benchmarks, enabling policy development and testing across MuJoCo, Isaac, and ManiSkill. The project facilitates zero-shot manipulation and accelerates robot learning with tools for data generation, grasp generation, and evaluation.

How It Works

The ecosystem provides a unified platform with diverse simulation assets (scenes, objects, robots) and an experiment configuration system for data generation and evaluation. Key features include scene conversion, automated grasp generation, and teleoperation. Assets are designed for interoperability across major simulators, with recent updates introducing benchmarks and leaderboards for standardized evaluation.

Quick Start & Requirements

Installation requires Python 3.11 (conda create -n mlspaces python=3.11). Clone the repo and run pip install -e ".[mujoco]". Optional installs include mujoco-filament or curobo (requiring CUDA 12.8 and specific PyTorch). Supports Linux/Mac. Data generation/benchmarking are MuJoCo-only. Simulator-specific asset setup instructions are linked. iPhone teleoperation requires the TeleDex app.

Highlighted Details

• Assets compatible with MuJoCo, Isaac, and ManiSkill.
• Large-scale datasets: ~129k Objaverse objects, ~120k procedurally generated scenes.
• Two benchmark versions (molmospaces_bench_v1, molmo_spaces_bench_v2) for task evaluation.
• iPhone teleoperation via TeleDex app.
• Recent updates (March 2026) include leaderboards, data generation code, and evaluation tools.

Maintenance & Community

Actively developed with recent updates in March 2026. Key contributors are listed in the citation. Community channels (Discord/Slack) are not explicitly mentioned.

Licensing & Compatibility

Codebase is Apache 2.0. Data licenses vary: Objaverse subsets are ODC-BY 1.0, others CC BY 4.0. Artifacts are intended for research/educational use under AI2's Responsible Use Guidelines, potentially restricting commercial applications.

Limitations & Caveats

Data generation and benchmarking are MuJoCo-only. Teleoperation is iPhone-restricted. Curobo installation is complex, requiring specific CUDA/PyTorch. Artifacts are designated for research/educational purposes only.