Real Robot Competition: AiRLIHockey - Air-HocKIT

News

[2023-12-19] The papers presenting the submitted solutions are available below on this page.
[2023-11-03] The workshop schedule at NeuRIPS is available here.
[2023-11-03] We're honored to have Christopher G. Atkeson and Sergey Levine as our invited speakers!
[2023-11-03] The second-round competitions are finished! We thank all participants for developing fantastic agents.
[2023-10-16] The first-round tournament has been finished! Please check the leaderboard for the videos and results!
[2023-09-12] The one-page reports of the qualified teams are now available here.
[2023-08-10] The Qualifying Stage will end on 11th August. The tournament stage will open on 14th Aug.
[2023-07-13] The Q&A session will be organized on July 20. 5:00-6:30 pm CET. Click to Join the Zoom Meeting!
[2023-07-13] A short video demo of the best-performing teams is now available here.
[2023-06-04] The Robot Air Hockey Challenge is now part of the NeuRIPS Competition Track!

Motivation

Closing the reality gap is one of the most important research directions toward true embodied intelligence. To reach this objective and deploy learning algorithms on a robot operating in the real world, we need to consider various practical issues: disturbances, observation noise, safety, model mismatch, delays, actuator limitations, physical feasibility, limited real-world interactions, etc.

In addition, we also expect intelligent robots to be reactive in dynamic environments, achieving fast and agile movements and long-horizon planning capabilities.

The Robot Air Hockey Challenge provides a platform for researchers in the field of robot learning to interact with each other on a realistic robotic task. In this competition, teams will design and build their air hockey agents, competing against each other in different subtasks (in simulation) and finally in an entire game (both in simulation and in the real world).

In this challenge, participants should develop a robot agent that:

is robust to environmental disturbances, such as non-ideal physics models, observation noise, and data loss.
pushes the limits of the robot's capability. The robot should hit the puck as fast as possible while dealing with the actuators' limits.
accomplishes safe and reasonable behavior. The robot should keep the end-effectors on the table and stay within its boundaries.
performs dynamic and agile motions. The robot should be reactive to interact with a highly dynamic environment.
adapts to the sim-to-real gap with a limited amount of real system interactions.
finds both the low-level strategy to control the mechanical systems and the high-level strategy to win the game.

Challenge Structure

The main part of the challenge consists of three simulated stages: Warm Up, Qualifying, and Tournament. Each stage consists of different tasks required for robot air hockey. Besides preparing your agents for executing the desired movements in the environments that are provided to you, they should also be robust and capable of adapting to changes. Except for the warm-up stage, we will evaluate your submissions in modified simulation environments to mimic the sim-to-real gap. Finally, the first three teams will be able to deploy their approach on the real robot.

In this challenge, the only requirement is that your solutions have some learning algorithm for one or more components of your agent. The agents should thereby adapt their behavior through interactions with the modified evaluation environments.

At the end of the qualifying and tournament phases, teams are required to submit a two-page report of the approach, which is open to the public.

More information can be found in the Challenge Details.