ICML Paper: Empowerment-Based Representations Are 1000x More Sensitive to Control-Relevant Features
burny_tech · x · 2026-10-25
A new arXiv paper from Benjamin Eysenbach's group, "Learning to Perceive the World Through Control," proposes learning representations via the empowerment objective — maximizing mutual information between an agent's skills and their future outcomes — instead of prediction goals, external rewards, or fixed datasets.
Key points:
- Empowerment agents induce two complementary representations: forward (which outcomes can be influenced) and backward (how states can be reached).
- Under specified assumptions, both become invariant to control-irrelevant information without task rewards or expert demonstrations.
- In Point-Maze, Ant, and visual manipulation experiments, the learned representations show at least 1000x greater sensitivity to control-relevant state variables than irrelevant ones.
- The authors argue that learning representations through interaction aimed at maximizing control — rather than from passive datasets — is essential for useful invariances, aligning with the causal learning literature.
More from Embodied
- Apollo Hospitals deploys Intuitive's da Vinci 5 surgical robot in Bengaluru — VectorWireAI · 2026-10-25
- ETH Zurich teaches G1 robot to climb ladders with just 10 minutes of human data — Distinct-Question-16 · 2026-10-25
- New soft actuator weighs 15g, lifts 400g, costs under $3 to build — IanPritchard · 2026-10-25
- Mannequin rigged with robot arms seen hunched over a soldering bench — repligate · 2026-10-25
- NVIDIA Open-Sources 2,575 SimReady 3D Assets, 472GB With Physics Data — tinyfool · 2026-10-25
- Robotics researcher: 2 days is enough to ship a follow-up paper on supermimic — ChongZzZhang · 2026-10-25