Pages 1–5 of the second lecture distinguish training loss from evaluation and compare how Absolute Zero, R-Zero, SPICE and R-Few construct self-training loops.
Topic: #reinforcement-learning
A collection of 20 posts about reinforcement-learning.
Proposer–solver self-play connects R-Zero’s plateau, execution checks in Absolute Zero and external information in SPICE and R-Few.
Which part of a system changes, and what feedback drives it? Evaluation and generalization evidence put claims about recursive self-improvement in context.
Worked examples contrast revising an answer with updating parameters through supervised learning on corrections or reinforcement learning on revision actions.
A counterexample shows how intermediate rewards can change the objective. Potential-based shaping, generated reward code and policy evaluation offer ways to check the effect.
Eureka, REvolve and RF-Agent introduce reward search. The lesson connects dopamine prediction errors, preference-based reward models, RLHF and RLAIF.
How ECoT and inference-time scaling let robots spend computation on decisions.
Turning generalization, contact sensing, rapid adaptation, and lifelong learning into research questions.
How one model can control multiple robots.
How world models predict the consequences of robot actions.
From rewards and values to Bellman equations, value and policy iteration, Q-learning, DQN and continuous-action control. Compare the original figures with timestamped transcripts and work through small grids and numerical examples.
A guided reading of REINFORCE, baselines and advantages, importance sampling, TRPO, PPO, and SAC, checked against the original slides and public captions. Worked calculations clarify probability ratios, entropy, and loss signs.
Why can averaging fail when a scene admits several correct actions? Explore VAEs, VQ-VAEs, DDPM and DDIM, Diffusion Policy, and flow matching through original figures, public captions, and supplementary lectures.
How can a robot remember states that one photograph cannot reveal? From attention, tokens, and position information to CLIP, VLMs, ACT, and FAST, this chapter compares slides with public captions and works through numerical examples.
What do robots learn, and how do they act? Follow the ETH 2026 lecture slides through policies, imitation learning, reinforcement learning and VLAs, then connect the ideas to working code and simulation examples.
Follow a robot's goal into physical motion through coordinate frames, inverse kinematics, trajectories and PID control, then build an MDP from states, actions, transitions and rewards. Read the ETH Lecture 2 slides alongside the transcript and a simulation I ran.
Read Ataraxos through the original Stratego board and game graphs. Understand hidden information, beliefs, self-play and the 15–1–4 result without confusing game performance with real-world decision-making.
JustGRPO, an ICML 2026 Outstanding Paper, connects diffusion language models, left-to-right generation, reward learning, and Pass@k. Original figures and public code show which improvements the experiments measured.
Follow a chip design through to manufacturable layouts and see where AlphaChip, commercial EDA software and OpenROAD help engineers.
A review of “Deep Reinforcement Learning for Robotics: A Survey of Real-World Successes.”