Robotics & Embodied AI
Robotic learning system can store knowledge from past tasks and reuse it for longer tasks, researchers report
In a 2025 paper in Nature Machine Intelligence, researchers describe a robotic system that keeps learning as tasks arrive one at a time and later reuses that knowledge for longer jobs. The system combines a growing knowledge space with language embeddings, which are numerical representations of word meaning. In the abstract, the researchers report that the embodied agent, a robot that acts in the physical world, accumulated knowledge from a continuous stream of one-time tasks and then applied the stored knowledge to multi-step real-world tasks.
Why is lifelong learning hard for a robot?
Many robot learning systems focus on one task at a time. Treating learning as a continuing process is harder because a robot has to avoid losing old skills while taking in new ones. This paper presents a framework designed to let knowledge grow over time.
How does the system hold on to old knowledge?
The framework stores what the robot learns in a knowledge space inspired by Bayesian non-parametric modeling, which lets the number of learned groups grow with experience. It also adds language embeddings so the robot can connect what a task says to what it has learned before. According to the abstract, this allows the embodied agent to keep knowledge from tasks presented only once and reapply pieces of that knowledge to longer tasks.
FAQ
What is reinforcement learning?
Reinforcement learning is training an agent by giving it rewards or penalties as it acts. Over many attempts, the agent adjusts its behavior to get more reward.
What are language embeddings?
Language embeddings are numerical representations of word meaning. They help the robot link what a task asks to past learned knowledge.
Does the abstract report a numeric benchmark?
No. The abstract describes the approach and reports that the agent can complete these tasks, but it does not give a score or a comparison table.
Limits and what's next
The abstract describes a framework and its reported behavior, not a numerical result. It does not state how many tasks, robots, or environments the researchers used, or which long-horizon tasks the robot attempted. This is one study, and the abstract reports no independent replication. The authors say the work may inspire the development of more broadly applicable intelligence, so future work would need to show whether the approach holds up in new settings.
Source: Yuan Meng et al., “Preserving and combining knowledge in robotic lifelong reinforcement learning,” Nature Machine Intelligence, 2025.