First method to estimate empowerment from offline data for intrinsically motivated exploration — Berkeley/Princeton
ben_eysenbach · x · 2026-11-05
Ishir Garg (UC Berkeley), with Benjamin Eysenbach (Princeton), Sergey Levine, and collaborators, propose the first method for estimating empowerment from unlabeled offline data — no environment interaction required.
- Context: a common agentic recipe trains a policy on prior data then improves it with online rollouts, but little work explores how prior data can guide the online exploration phase
- High-empowerment states require more exploration (more possible futures), yet existing estimators need online policy rollouts
- Method: jointly learn an empowerment-maximizing policy and its discounted state-occupancy measure, trainable on unlabeled trajectories from any unknown policy
- The estimator yields intuitively sensible empowerment measures and improves exploration in online goal-conditioned RL
- Paper, code, and project page are public
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