Robot learning

Learned policies (ACT, Diffusion Policy, OpenVLA, π0, SmolVLA), the benchmarks they are measured on, and the language models robots now call for planning and judging. It also holds the older ideas about uncertainty (ensembles, calibration, abstention) that I keep returning to.

Concepts

• Embodied AI benchmarks: what they measure and what they leave out.
• Evaluating language models: how LLM judges fail.
• Byzantine-robust learning: training when some workers lie.
• Disagreement between models, filed under perception.

Entities

Datasets, platforms, tools and groups that come up in this subfield: GPT-4, Gemini, Habitat, Gazebo, Franka arms. The entity index has the rest.

Open questions

• What would a benchmark look like with three robots, one of them broken, and a task that lasts a week?
• Can a data generator randomise faults and attacks as easily as textures?
• A robot that asks a VLM whether the grasp worked is using an LLM judge. How often does it hear yes when it shouldn’t?

Papers

Other subfields: perception, mapping and localisation, semantic mapping, planning and navigation, control, multi-robot systems, safety and verification, security, Human-robot interaction, field robotics.