Third-person imitation learning capability added to ML platform

AI Impact Summary

The change enables training policies from third-person demonstrations (observed tasks performed by other agents) rather than only first-person trajectories. This broadens data sources and can accelerate learning for robotics, embodied AI, or simulation tasks, potentially lowering data collection costs and improving generalization across agents. Implementation considerations include adding an observation-enabled imitation pipeline (video/trajectory ingestion, pose/state extraction, cross-agent alignment) and adjusting reward models (inverse RL or behavioral cloning variants) to cope with off-policy demonstrations and domain gaps. Rigorous evaluation will be needed to guard against bias from demonstrator behavior and ensure policy robustness across environments.