Google DeepMind Achieving Human Level Competitive Robot Table Tennis

Method overview.

Overview of the Robot Table Tennis Agent

  • The research presents a robot agent capable of playing competitive table tennis at an amateur human level.
  • The robot achieved a 45% win rate against human opponents of varying skill levels in 29 matches.
  • Key contributions include a hierarchical policy architecture, techniques for zero-shot sim-to-real transfer, and real-time adaptation to opponents.

Contributions of the Study

  • The architecture consists of low-level controllers (LLCs) for specific skills (e.g., forehand topspin, backhand targeting) and a high-level controller (HLC) that selects the appropriate LLC based on game context.
  • Each LLC is trained to handle different aspects of table tennis, allowing for efficient skill specialization and adaptation.
  • Hierarchical and Modular Policy Architecture Low-level controllers with skill descriptors to model agent capabilities. High-level controller for selecting low-level skills.
  • Zero-shot Sim-to-Real Techniques Iterative task distribution approach grounded in real-world scenarios. Automatic curriculum definition for training.
  • Real-Time Adaptation Ability to adapt to unseen opponents during matches.

Training Methodology

  • A hybrid training approach combines reinforcement learning (RL) and imitation learning (IL) to develop the robot's skills.
  • Initial training data is collected from human-human play, which is then used to create a realistic task distribution for simulation training.
  • The robot iteratively refines its skills through a cycle of simulated training and real-world evaluation, enhancing its performance over time.


Real-Time Adaptation and Performance Assessment

  • The robot adapts to unseen human opponents by tracking match statistics and adjusting its strategy in real-time.

  • Conducted 29 matches between the robot and human players.
  • Robot won 45% of the matches (13 out of 29).
  • Performance breakdown: 100% win rate against beginner players. 55% win rate against intermediate players. Lost all matches against advanced players.
  • Participants reported a positive experience, finding the robot engaging and fun to play against.

?Limitations and Future Work

  • The robot struggles with fast balls, low balls, and high spins, indicating areas for improvement in its capabilities.

Source: https://huggingface.co/papers/2408.03906


要查看或添加评论,请登录

Srinivas Hebbar的更多文章

社区洞察

其他会员也浏览了