The Challenge
The goal of this year’s challenge is to solve dexterous manipulation tasks with offline reinforcement learning (RL) or imitation learning. The participants are provided with datasets containing dozens of hours of robotic data and can evaluate their policies remotely on a cluster of real TriFinger robots.
Participants can tackle two tasks during the real-robot stage:
- Pushing a cube to a target location on the ground and
- lifting the cube to match a target position and orientation in the air.
Like last year’s challenge, the Real Robot Challenge III is featured in the NeurIPS 2022 Competition Track.
Motivation
Learning manipulation policies on real robots is usually quite costly and challenging. Offline RL, however, has the potential to mitigate this problem by making use of precollected data. Yet, opportunities for researchers to work with such datasets and to then evaluate learned policies on real-robot hardware are scarce. For this reason, a large part of the RL community uses simulators to develop and benchmark algorithms. However, insights gained in simulation do not necessarily translate to real robots, in particular for tasks involving complex interactions with the environment. The purpose of this competition is to alleviate this problem by allowing participants to experiment remotely with a real robot -- as easily as in simulation.
We hope that the Real Robot Challenge series will
- lead to a more inclusive and coordinated robotic learning community,
- allow for the generation of orders of magnitude more data than what is possible in individual labs with typical robotic platforms and
- encourage the development of offline RL algorithms which can make efficient use of real-world data.
These factors may help advancing the state of the art in dexterous manipulation, which has myriad of applications, such as construction, cleaning, care of the sick and elderly, agriculture and assembly.
