Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models

Researchers introduce Embodied-R1.5, an 8-billion-parameter foundation model that achieves state-of-the-art performance on embodied AI tasks by integrating reasoning, planning, and self-correction capabilities. The model demonstrates strong generalization to real-world robotics applications and is being open-sourced with training code and evaluation tools.

Embodied-R1.5 represents a significant advancement in physical AI systems by consolidating multiple embodied reasoning capabilities into a single, efficient architecture. The model's architecture centers on a Planner-Grounder-Corrector framework that enables autonomous task execution and self-correction over extended sequences, addressing a critical limitation of previous systems that struggled with error recovery and multi-step reasoning in dynamic environments.

This development emerges from years of progress in embodied AI, where the field has gradually shifted from task-specific systems to foundation models capable of generalizing across domains. The construction of a 15-billion-token dataset through automated pipelines represents substantial technical progress in scaling embodied AI training data, a historically limiting constraint. The multi-task balanced reinforcement learning recipe addresses the fundamental challenge of training unified models across heterogeneous robotics tasks that often conflict during optimization.

The practical impact extends beyond benchmark improvements. By achieving state-of-the-art results on 16 of 24 embodied vision-language model benchmarks while maintaining only 8B parameters, Embodied-R1.5 establishes efficiency standards for the field. The model's ability to fine-tune into vision-language action (VLA) systems with minimal additional data suggests a path toward more accessible robotics development for smaller organizations and research teams. Real-world validation across instruction following, affordance grounding, and complex manipulation tasks demonstrates genuine progress toward deployable systems.

The open-sourcing of weights, datasets, and the EmbodiedEvalKit evaluation framework signals a shift toward democratized embodied AI research. This approach could accelerate innovation by enabling broader participation and establishing standardized evaluation methodologies. Watch for rapid iteration cycles as the community builds upon these foundations and whether the efficiency gains translate to commercial robotics applications requiring real-time inference.

• →Embodied-R1.5 achieves state-of-the-art on 16 of 24 embodied AI benchmarks with only 8B parameters using a unified foundation model architecture
• →The model integrates an autonomous Planner-Grounder-Corrector framework enabling self-correction for long-horizon robotic tasks
• →A 15-billion-token dataset constructed through automated pipelines expands training data availability for embodied AI tasks significantly
• →The system demonstrates strong real-world generalization across instruction following, affordance grounding, and complex manipulation tasks
• →Full open-sourcing of model weights, code, and evaluation tools enables broader participation in embodied AI research development