Gongumenn

Our research demonstrates progress in training robots to play table tennis against human opponents. However, several
limitations warrant further investigation: (1) Fast Balls: The robot struggles to react effectively to very fast balls due
to multiple factors; lack of data, high system latency, and between-shot reset. Even with a high control frequency
(50hz), the inherent latency in the system (∼100ms) significantly restricts the number of actionable decisions the
robot can make within a given timeframe. This problem was exacerbated by the fact that the robot was reset to
the same initial pose — close to the table, paddle facing forward — in between each ball hit to account for the single
ball training. While increasing the control frequency could mitigate this issue, it reduces the available time per action for
computation. (2) High Balls: The robot is unable to handle balls higher than ∼6f t above the table, as it goes outside the
camera FOV. In practice, it was seldom a problem. (3) Low Balls: The robot’s inability to consistently handle very low
balls is primarily attributed to collision avoidance protocols built into the system. These protocols, while essential for
protecting the robot’s paddle, pose a significant challenge in handling balls that stay close to the table. This limitation was
frequently exploited by skilled players by sending underspin balls that stay very low. (4) Short Pips Paddle: In this work,
we chose a paddle that can impart spin, while still being possible to model within the simulator. The paddle doesn’t
have the smooth rubber found on advanced table tennis paddles [17]. We found accurately modeling such advanced
paddles extremely challenging given the highly nonlinear and multimodal nature of the paddle. This limitation was frequently noted by human participants and highlights the need
for more advanced paddle modeling techniques. (5) Spin Detection The policy is unable to accurately read extreme spins on an incoming ball. This limited the amount of spin that
could be effectively handled, a fact that was exploited by a number of opponents. It was possible to train specialist LLC
policies to handle different types of spin, so improvements in spin detection would likely directly translate to improved
overall performance. (6) No Multi-ball Strategic Game Play: The robot’s gameplay is predominantly ”one ball at a
time”, lacking the multi-ball strategic play observed in skilled human players. This limitation is attributed to the single-ball
training approach. (7) Motion Capture Unreliability: The reliance on motion capture technology introduces a degree of
unreliability, as occasional glitches or inaccuracies in motion tracking can negatively impact the robot’s performance. (8)
Backhand The policy had weaker backhand play compared with the forehand. (9) Predictability The policy did vary
its placement however this could be improved further. Being less predictable in placement of certain balls was one of the
top three areas for improvement as identified by the match referee (in addition to reading spin better and improving the backhand).