Stanford launches 2026 BEHAVIOR Challenge

The 2026 BEHAVIOR Challenge highlights a major shift in Embodied AI evaluation by forcing models to operate strictly on onboard sensor data (no external tracking) while tackling longer and more complex bimanual tasks. By doubling both the tasks and human demonstration data, Stanford is setting a highly realistic standard that will reveal exactly how well current foundation models can generalize to noisy, real-world household robotics.

• –**Strict Onboard Tracks**: Limiting the official evaluation track to RGB, Depth, and Proprioception mirrors real-world deployment challenges where external cameras/systems are not available.
• –**Long-Horizon Scaling**: At an average of 6 minutes per task, models must exhibit high levels of planning, spatial memory, and error recovery rather than just executing reactive motor skills.
• –**Improved Baselines**: Providing support for modern baselines like pi0.5 and GR00T N1.7 makes it easier for research teams to start and benchmark their own architectures.