Chapter 3: Training

Core Questions

• Can foundation models serve as zero-shot planners?
• How do we bridge the gap between token prediction and physical action?
• What role does chain-of-thought play in robotics?

Topics

2.1 Planning vs. Execution

• The inference-to-control pipeline
• Temporal credit assignment
• Failure recovery strategies

2.2 Foundation Models as World Models

• Zero-shot task adaptation
• In-context learning for robotics
• The limits of pre-training

2.3 Chain-of-Thought in Physical Space

• Decomposing complex tasks
• Spatial reasoning with LLMs
• Error propagation in multi-step plans