The world of chess and robotics collide in a fascinating event that showcases the challenges of translating complex strategies into physical action. IIT Delhi's RoboGambit competition is a testament to the intricate relationship between artificial intelligence and the tangible world. This article delves into the unique experience of witnessing chess engines struggle with the limitations of their robotic arms, offering a fresh perspective on the capabilities and limitations of AI in chess.
The competition, organized by Nitye Gupta and Sanidhya Ojha, brought together 15 hostel teams from the main campus and two from the Abu Dhabi campus. The rules were a modified version of traditional chess, played on a 6x6 board without rooks or castling. Pawns could only move one square forward, and promotion was restricted to pieces that had already been captured and were off the board. This variant, never played before, presented a unique challenge for the teams.
The software simulation phase was a hurdle for many, with only 11 teams successfully building a working chess engine. The hardware round, however, proved even more demanding, as only 8 teams managed to integrate their software with the physical robot arms. This highlighted the difficulty of translating theoretical chess strategies into practical, timed movements.
The teams had to teach their engines to understand a game that no one had ever played before. Modern chess engines rely on two core functions: search and evaluation. The search function explores future moves, while the evaluation function assesses the quality of positions. Traditionally, evaluation was done by hand, with programmers encoding rules into the software. However, the breakthrough came with the use of neural networks (NNUE) that mimic human intuition, leading to superhuman performance in engines like Stockfish.
In the RoboGambit competition, teams had to build their evaluation networks from scratch. The team Magnus Galgotias, from Nilgiri Hostel, initially attempted reinforcement learning, but it proved too resource-intensive for the limited time. They switched to supervised learning, manufacturing their own training data by playing 50,000 games against themselves. This approach, combined with a traditional search engine built in C++, resulted in an engine playing at a 2100-rated chess.com level, an impressive feat in just a week.
The real challenge, however, lay in the physical execution of the game. The competition was a relay between humans and robots, with a shared ten-minute clock for each move. The human players had to find moves that the robotic arm could execute quickly and cleanly, all while managing the clock. Captures, in particular, were a hurdle, requiring the arm to lift and set aside pieces, adding complexity to each turn.
The top four teams, Girnar, Satpura, Nilgiri, and Kailash, showcased their prowess in both software and hardware. Girnar's victory was attributed to their ability to handle the robotic arm effectively, while Satpura's engine made a critical error, leading to a loss. Nilgiri's close encounter with Satpura highlighted the importance of mechanical dexterity and composure, as they lost due to time constraints despite having a strong position.
RoboGambit offers a unique insight into the intersection of AI and chess. It demonstrates the challenges of translating complex strategies into physical actions and the importance of mechanical precision and composure. The competition serves as a reminder that while AI has made remarkable strides, there is still a long way to go in bridging the gap between theoretical chess and practical, timed gameplay.
