We can learn a model that provides shaped "process rewards" for robotic RL, that evolves automatically as the policy gets better. This improves performance on benchmarks, and works in the real world! Some fun new work with Raymond Tsao & @ajwagenmaker
Users praise the new method for learning evolving shaped rewards in robotic reinforcement learning as cool because dense rewards remain highly effective for task mastery.
No Digg Deeper questions have been answered for this story yet.
In the real world, this leads to drastically faster learning, in just a few thousand time steps.
We can learn a model that provides shaped "process rewards" for robotic RL, that evolves automatically as the policy gets better. This improves performance on benchmarks, and works in the real world! Some fun new work with Raymond Tsao & @ajwagenmaker
To learn more, check out website: https://success-visitation-matching.github.io/ paper: https://arxiv.org/abs/2606.23640
In the real world, this leads to drastically faster learning, in just a few thousand time steps.
@svlevine @ajwagenmaker This is cool! Every time I do robotics research I remember dense reward is still overpowered for task mastery, and the more self supervised we can learn the better :)
