Robot Learning 2026 – Lecture 10: Embodied Reasoning and Test-time Scaling | ETH Zürich

Intelligent reasoning serves as a critical frontier in robot learning, enabling models to generalize beyond training distributions by treating reasoning as a first-class component of policy design. Embodied Chain of Thought architectures improve performance by 30% by generating intermediate plans and grounded reasoning steps before action execution. While this process introduces latency, techniques like reasoning dropout and decoupled pre-training allow for internalizing these representations to maintain real-time control. Furthermore, test-time compute scaling—leveraging reinforcement learning algorithms like GRPO—allows models to solve complex, novel problems by systematically amplifying reasoning capabilities through verifiable reward signals. This approach shifts the focus from merely collecting massive datasets to optimizing the model’s internal logic, enabling robots to adapt and perform effectively in out-of-distribution scenarios without requiring additional human-annotated demonstrations.