07/30/2026
Excited to share our new work: "SymmGrid — Super-Scaling On-Robot Learning with Parallelized Symmetries and Egocentric–Exocentric Visual Perception".
One of the major barriers to reinforcement learning directly on physical robots is simple: it takes too long.
With SymmGrid, we ask whether we can extract substantially more learning value from every real-world robot interaction.
The idea is surprisingly simple: use parallelized branched symmetries to transform each physical trajectory into many valid, symmetry-consistent experiences. These equivalences enrich the replay buffer without requiring the robot to physically execute every transformed trajectory.
We also address a practical challenge that becomes important when moving from simulation to real robots: perception. SymmGrid supports both egocentric and exocentric cameras, using efficient homography-based visual transformations for fixed external cameras.
We evaluated SymmGrid directly on a Franka FR3 across three contact-rich manipulation tasks:
🔹 Peg insertion: 100% evaluation success in as little as 16.6 min
🔹 Cable routing: 98% evaluation success in 10.9 min
🔹 Object relocation: 92% training success in as little as 79.3 min
Across the experiments, SymmGrid achieved 1.37–2.17× faster wall-clock convergence than the SOTA baseline and improvements of up to 2.59× in normalized area under the learning curve (nAUC).
Perhaps what excites me most is where this points: real-robot reinforcement learning approaching the 10-minute range for contact-rich manipulation.
Faster on-robot learning means relying less on approximations of real contact physics in simulation—and moving closer to robots and humanoids that can efficiently learn new physical skills directly in the real world.
I’m grateful to have worked with a team of exclusively undergraduate students on this project as part of our lippyRobotics undergraduate research lab: Gabe Everett, Brice Gunter, Ryan Vander Stelt, Cleiver Ruiz-Martinez, and Blake Hull.
It's no small thing to compete at this level of research and performance with only undergraduate students.
See our project page for
📄 Paper | 💻 Code | 🎥 Videos | 🤗 Hugging Face | 📚 Tutorials:
https://lnkd.in/g89UcKtn
symmetries
