“Teach a Robot to FISH” Wins Best Student Paper Award at 2023 Robotics Science and Systems Conference

July 27, 2023

“Teach a Robot to FISH: Versatile Imitation from One Minute of Demonstrations" won the Best Student Paper Award at the 19th “Robotics: Science and Systems” (RSS) Conference hosted in Daegu, Korea earlier this month. The paper was co-authored by Siddhant Haldar (a second-year doctoral student), Jyothish Pari (a recent B.A. graduate in Mathematics and Computer Science), Anant Rai (a recent M.S. graduate in Computer Science), and Lerrel Pinto (Assistant Professor of Computer Science). With co-authors at the undergraduate, graduate, doctoral and faculty levels, the award-winning paper demonstrates the impressive collaboration among Courant researchers.

The team sought to confront the notoriously difficult challenge of teaching robots new tasks. “Learning skills through real-world interactions is hard for robots,” say Siddhant Haldar and Jyothish Pari, the primary authors of the paper, “reinforcement learning often requires a large amount of time or a human operator to be omnipresent.” Hoping to save both time and energy, the researchers developed FISH: Fast Imitation of Skills from Humans. “In our work, we propose an algorithm for efficient robot learning that requires as little as 30 minutes to teach a new task,” they say. 

How do you train a robot to perform a new task quickly and efficiently? You teach it to learn from its mistakes. “FISH computes rewards that correspond to the ‘match’ between the robot’s behavior and the demonstrations,” the paper describes, “these rewards are then used to adaptively update a residual policy that adds on to the base policy.” The authors utilized FISH to test robots on their ability to learn household tasks such as inserting a key into a lock, opening a door, and flipping bread in a pan—this last test was a success but it resulted in “a table full of rock-hard stale bagels,” Haldar and Pari admit.

The research team often conducted tests separately and discussed their findings afterwards. Haldar and Pari explain, “We divided the robotics experiments, and whenever one of us would improve the performance on a specific robot, we made sure that the algorithmic improvement was beneficial across all the different robots.” In the final month before their paper’s submission, the authors would stay late at the lab running experiments. They devised a nightly ritual for sustaining their energy: “We would all gather around 10pm every night, make ourselves a cup of tea, and play a couple matches of FIFA before getting back to work again.”

The RSS Best Paper Award proves their hard work—occasionally punctuated by video games—has paid off. The initial results are promising. On average, the new algorithm achieves a success rate of 93% which is about 3.8 times higher than previous state-of-the-art-methods. In Haldar and Pari’s view, there is still room for improvement. “This is just the initial step in this direction,” they say, “in the future, we need to develop both efficient algorithms for real-world learning and robust hardware capable of real-world interactions.” This goal is reflected in the title of the paper—if you can teach a robot to FISH, you may keep it occupied for a lifetime.