EmbodiedBrain: Expanding Performance Boundaries of Task Planning for Embodied Intelligence

We present EmbodiedBrain, a large-scale embodied planning model trained with multimodal post-training and reinforcement learning.

EmbodiedBrain is built upon Qwen2.5VL-7B and 32B backbones and targets long-horizon task planning for embodied intelligence. We construct large-scale post-training datasets with 235k SFT samples and 118k RL samples, and propose Step-GRPO, an improved policy optimization algorithm tailored for long-horizon planning tasks.

A comprehensive evaluation framework is established, covering:

• Multimodal general capability benchmarks
• Static embodied planning benchmarks
• Simulation-based evaluation with AI2Thor

EmbodiedBrain significantly outperforms RoboBrain2.0 across multimodal reasoning, spatial perception, and task planning benchmarks.

The technical report and codebase are publicly released.