@GTrobotics
🚀 Can we do even better than DAGGER (a popular imitation learning algorithm in RL) in terms of sample efficiency? Absolutely!
Check out our latest work on Learn to Teach (L2T)—a one-stage training framework that seamlessly unifies teacher and student policy learning. This work is led by our ML Ph.D. student @feiyangwu and his team. By recycling simulator samples and synchronizing learning trajectories through shared dynamics, L2T slashes sample complexity and training time while achieving the SOTA performance.
But we didn’t stop there. We put L2T to the test with extensive simulations and hardware trials on our Digit robot, proving its zero-shot sim-to-real transfer across 12+ tough terrains—all without depth estimation modules. The result? Unmatched robustness and adaptability in real-world scenarios.
If you're into RL for robotics, this is one you don’t want to miss! 🚀🤖 @MEGeorgiaTech @GTrobotics @mlatgt
arxiv.org/pdf/2402.06783…
lidar-learn-to-teach.github.io
#ReinforcementLearning #AI #Robotics #MachineLearning #SimToReal #RLView quoted post on X
