RPL

At the very early stage of RPL (amazon-far/RPL), @younggyo and I discussed a similar idea during multi-terrain rl training — using terrain-level as a proxy of task success rate to guide terrain configuration sampling at environment resets. It indeed improved the policy performance on multiple terrains, though our two-stage distillation pipeline ultimately worked better in our experiments, at the cost of separate expert training. Nice to see success-guided sampling studied systematically and at this scale!

RPL code is released at amazon-far/RPL, including experts training, multi-depth distillation and sim2sim. If you have an Xbox controller lying around, it’s a lot of fun playing it in Mujoco. Its perceptive features were adopted by some cocurrent work like LadderMan (ladderman-robot.github.io) and PHP (php-parkour.github.io) with promising results. RPL was also accepted to CoRL2026. See you in Austin.