ChronoSRL Gives RL Critics Temporal Geometry, Outperforming CRL and SRL Across Network Depths
Jan_R_Peters · x · 2026-10-01
Researchers Nico Bohlinger and Jan Peters (TU Darmstadt) introduce ChronoSRL, which adds temporal geometry to the critic of self-supervised RL: distances in the critic's embedding space are trained to match the time an agent needs to reach a goal (distance ≈ κτ), with heads predicting when a goal is reached and how long the agent stays.
Key idea: a goal close in space can be far away in time, so embedding distance should measure time-to-goal, not spatial proximity. Never-reached goals are pushed beyond κW.
Results:
- Compared against CRL, AC-CRL and SRL on 7 tasks across network depths from 1 to 64
- Faster learning (AUC) and higher final performance, even with much smaller networks
- In Ant Hardest Maze, its distance tracks remaining steps more closely than spatial distance
Demonstrated on sim-to-real quadruped locomotion with a Unitree Go2 (velocity tracking, goal reaching, box climbing). Paper, code and an online playground are available.
More from Embodied
- Most pure robotics deployment firms aren't profitable, says founder who's talked to many — ATTlKA · 2026-10-01
- Pollen Robotics: 10,950 Microducks to ship between December 10 and January 10 — andimarafioti · 2026-10-01
- OpenAI and Meta bet on cute software agents to soft-launch AI hardware — haydenfield · 2026-10-01
- Runway unveils Praxis-1, an open-weight World Action Model for robot control from video — runwayml · 2026-10-01
- QCQP-representable IMU pre-integration factor brings certifiable estimation to inertial SLAM — zhenjun_zhao · 2026-10-01
- Pow3R-SLAM runs 1.6x faster than MASt3R-SLAM with 15% lower trajectory error — zhenjun_zhao · 2026-10-01