GCRL and unsupervised skill learning unified via control-maximization — NeurIPS Oral from Eysenbach and Dayan
ben_eysenbach · x · 2026-11-05
A NeurIPS Oral paper by Modirshanechi, Benjamin Eysenbach, Peter Dayan, and Eric Schulz unifies goal-conditioned RL (GCRL) and mutual-information skill learning (MISL) as instances of control maximization.
- Motivation: why skills learned via MISL support downstream goal-reaching lacked theoretical grounding; both GCRL and MISL are umbrella terms with many variants
- Identifies three canonical GCRL formulations (persistent goal, exact timing, opportunity window) and proves they are fundamentally inequivalent — they can induce incompatible optimal policies in the same environment
- Common interpretation: a good goal-conditioned policy makes future trajectories highly sensitive to the commanded goal; MISL objectives are measures of skill sensitivity akin to goal sensitivity
- Establishes a correspondence between GCRL formulations and MISL objectives, giving the first theoretical account of unsupervised pretraining's empirical success in GCRL
More from Embodied
- Open-source robotics talk at Wawtech Warsaw aims to get software devs building robots — kamathsblog · 2026-11-05
- PhD student blogs on challenges of latent action models for robotics pretraining — EugeneVinitsky · 2026-11-05
- Voxel-based sim2real author flags two hurdles: voxel resolution vs speed and data scarcity — chris_j_paxton · 2026-11-05
- Liquidating ~18 humanoid robots in SF: Unitree G1/H1/R1, AGIbot, Booster, EngineAI in stock — cixliv · 2026-11-05
- Nothing CEO says mysterious AI-first device launches next month — TechCrunch AI · 2026-11-05
- Pretraining Makes Robot Policies Reliable but Hurts Posttraining, New Paper Finds — siddkaramcheti · 2026-11-05