Robotics & Embodied AI
Robots learned to avoid collisions in crowded, complex scenes using sensor readings instead of constant communication

In a 2020 paper in The International Journal of Robotics Research, researchers from the University of Hong Kong, Baidu, and Fuzhou University describe a decentralized policy that turns each robot's raw sensor measurements into steering commands for movement velocity. The system was trained in simulation on many robots at once, then tested in simulated and real-world settings. The tests included a group of robots of different kinds, a scene with 100 robots, and one robot moving through a dense human crowd. The learned policy also worked on physical robots whose shapes and dynamics differed from the simulated agents.
What is different about this approach?
Many earlier distributed systems needed frequent communication between robots or special agent-level features to choose a safe local action. This policy instead uses raw sensor data and turns it directly into steering commands. The authors describe it as a decentralized sensor-level policy that maps raw sensor measurements directly to steering commands.
How was it tested?
The researchers trained the policy with a reinforcement-learning algorithm across many robots in complex simulated environments. They then used a hybrid control framework to improve its reliability and effectiveness, and evaluated it in simulated and real-world scenarios. In unseen scenes, they checked generalization with a group of robots of different kinds and a large scene with 100 robots. They also deployed it on physical robots with different shapes and dynamics from the simulated agents. The same policy guided a single robot through a dense real human crowd and helped it keep moving without getting stuck.
Frequently asked questions
Did the robots need to communicate with each other?
The abstract does not describe inter-robot communication. Each robot decided its own steering action from sensor measurements, with limited observation of other robots. The abstract notes prior systems often required frequent inter-robot communication, whereas this policy acts on each robot's own sensor measurements.
Did it work outside simulation?
The policy was trained in simulation only, but the researchers ran it on physical robots with different shapes and dynamics. It also guided one robot through a dense real human crowd without getting stuck.
What does decentralized mean here?
It means each robot chooses its own collision-avoidance action from limited local observations, rather than a central planner telling every robot what to do.
Limits and what's next
This is one study, and the abstract reports no independent replication. The training used simulation data only, so the real-world results depend on the transfer the authors describe rather than on training that included physical-robot data. The evaluation covered simulated and real-world scenes, including a group of robots of different kinds, a scene with 100 robots, and one robot in a dense human crowd. The abstract does not report numerical performance scores for these scenarios. The authors frame the work as a first step toward narrowing the performance gap between decentralized and centralized methods.
Source: Tingxiang Fan et al., “Distributed multi-robot collision avoidance via deep reinforcement learning for navigation in complex scenarios,” The International Journal of Robotics Research, 2020.