HMN 2025: How Robotic arm successfully learns 1,000 manipulation tasks in one day

Over the past decades, roboticists have introduced a wide range of systems that can effectively tackle some real-world problems. Most of these robots, however, often perform poorly on tasks that they were not trained on, particularly those that entail manipulating previously unseen objects or handling objects that were encountered before in new ways.

Researchers at the Robot Learning Lab at Imperial College London recently developed a new imitation learning approach that could allow robots to successfully learn new tasks faster and without requiring substantial training data. Using this method, which was introduced in a paper published in Science Robotics, they were able to train a robotic arm to complete 1,000 different tasks in a single day.

“The research was initially inspired by our prior work on trajectory transfer, where we introduced a method that proved robust and efficient for teaching robots single tasks,” Kamil Dreczkowski and Pietro Vitiello, co-authors of the paper, told Tech Xplore.

“Our supervisor, Edward Johns, then presented us with an ambitious, but exciting, goal: to teach a robot a very large number of tasks in minimal time, specifically ‘a thousand in a day.’ Given that our existing method required less than a minute per task and enabled immediate deployment without post-demonstration network training, we believed this goal was highly feasible by extending our method to a multi-task learning setting.”

To tackle the challenge posed to them by their supervisor, Johns, Dreczkowski and Vitiello first performed a systematic investigation of fundamental priors that impact the effectiveness of imitation learning. The goal of their study eventually became to assess the effectiveness of various design choices in allowing robots to reliably learn multiple tasks in a short period of time.

A new method to enhance imitation learning

Imitation learning entails teaching robots new skills by presenting them with human demonstrations, videos or other examples of humans tackling the tasks that the robot should complete. Despite their promise, most imitation learning techniques require extensive human demonstration data to achieve good results.

“Ultimately, the paper aimed to challenge the prevailing assumption that highly capable robotic systems require massive datasets, complex neural policies, and large models for learning at scale,” said Dreczkowski and Vitiello. “As part of our study, we investigated two priors for robot learning: trajectory decomposition and retrieval-based generalization.”

The first prior that the researchers incorporated in their approach, known as trajectory decomposition, entails the separation of each object manipulation trajectory into two phases. These phases are dubbed the alignment and interaction phase.

“During the alignment phase, the robot positions its end effector or a grasped object relative to a target object—think of a robot aligning a plug with a socket,” explained Dreczkowski and Vitiello.

“Then, during the interaction phase, the robot carries out the actual manipulation, for instance inserting the plug into the socket. These two phases have different requirements, so using two different policies sequentially achieves an order of magnitude of improvement in data efficiency in the low-demonstration-per-task regime.”

The second prior that the team’s approach is based on is retrieval-based generalization. In contrast with various other deep learning methods, which encode all behaviors into so-called network weights, their method stores all demonstrations in a memory component.

“At test time, this method uses the language description of a task and an observation of the environment to find the single most relevant demonstration from memory,” said Dreczkowski and Vitiello. “This demonstration is then used to inform the policy about how to align with the test object and how to interact with it.”

The team’s newly introduced approach, dubbed MT3 (multi-task trajectory transfer), is designed to retrieve and decompose demonstration data. Using a strategy known as retrieval, it finds demonstrations relevant to the task that the robot is asked to complete, and then adapts them to the target object and plans the robot’s motions, utilizing pose estimation and a planner component.

“For interaction, the method simply replays the demonstrated end effector motion,” said Dreczkowski and Vitiello. “The unique advantages of this method are that it is very efficient and interpretable. We found that it is very effective when demonstrations per task are limited, and because it relies on pose estimation, users can visualize what the robot will do before it executes the motion. Crucially, the robot is guaranteed to never do anything that was not explicitly demonstrated.”

Training robots on different tasks with little data

To test the potential of their imitation learning approach, the team used it to train the Sawyer robot, a robotic system comprised of a single arm, with an integrated gripper and camera. Notably, they were able to teach the robot to complete 1000 distinct manipulation tasks in under 24 hours, relying on a single demonstration per task. Notably, they were also able to deploy the policy immediately after each single demonstration, without the need for further training.

“We demonstrated that complex, large-scale robot learning is possible without massive datasets and large neural models,” said Dreczkowski and Vitiello. “To the best of our knowledge, ours is the first systematic, large-scale evaluation of multi-task learning in the few-demonstration-per-task regime, addressing a critical gap in the literature. Our approach provides a compelling, data-efficient alternative to the traditional monolithic behavioral cloning paradigm.”

The new imitation learning approach developed by the researchers could soon be improved further and implemented on other robotic platforms. Eventually, it could contribute to the introduction of robots that can reliably tackle a wider range of skills, without requiring extensive training data.

“MT3’s reliance on analytical steps makes its actions highly interpretable and trustworthy, providing a major advantage over the ‘black box’ nature of many deep learning methods,” said Dreczkowski and Vitiello.

“Moreover, what led to this very efficient method was a large-scale study of many different design choices. By comparing the performance of all these designs, when varying the dataset size and the diversity of tasks learned, we were able to gather several insights that could be useful to anyone looking to improve the efficiency of their robots in manipulating objects.”

As part of their next studies, Dreczkowski and Vitiello plan to continue developing data-efficient and effective robot learning strategies that can be instantly deployed on robots. Their goal is to reduce the human efforts and financial resources currently required to train robots.

“A key area for improvement will be moving away from the direct trajectory replay in our interaction phase,” added Dreczkowski and Vitiello. “We aim to explore methods for adapting trajectories to object geometries and changing test scenarios, allowing the robot to generalize more robustly to unseen variations that require more than a simple execution of the demonstrated motion.”

© 2025 Science X Network

The content is provided for information purposes only.