Dual Process Motion Planning
This paper proposes a neuro-symbolic dual-process motion planning framework that dynamically integrates fast, experience-driven learning ("System-1") with robust, symbolic reasoning ("System-2") to achieve superior efficiency, accuracy, and generalization in nonlinear robotic tasks.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Robots are moving from factory floors into our daily lives, tasked with everything from driving cars to navigating crowded hallways. To do this safely, a robot must solve a complex puzzle every second: how to move from where it is to where it needs to go without hitting anything, while obeying the laws of physics that govern its movement. For decades, engineers have relied on two main ways to solve this. One method is like a seasoned expert who calculates every possible path with extreme precision, ensuring safety but taking a long time to think. The other is like a fast instinct, where a robot uses past experiences to guess a good path instantly, but this guess might sometimes be wrong or unsafe. The challenge has always been finding a way to get the speed of the instinct without losing the safety of the calculation.
A team of researchers at the Chinese University of Hong Kong, Shenzhen, and the University of Oxford has proposed a new way to handle this trade-off. They built a system that mimics how humans often make decisions, combining a fast, intuitive reaction with a slower, more careful check. They call this a dual-process architecture. In their setup, the robot first tries to solve the problem using a fast, learned policy—a kind of digital intuition trained on thousands of successful past trips. This "System 1" acts quickly, proposing a path in a fraction of a second. However, before the robot actually moves, a safety gate checks if this fast guess is truly collision-free and actually reaches the goal. If the fast guess passes the check, the robot moves immediately. If the fast guess fails or looks risky, the system instantly switches to a slower, more rigorous calculator—a "System 2"—that works out a perfect path from scratch, just like the traditional expert methods.
The researchers tested this approach in a variety of difficult digital environments, ranging from open spaces with a few large obstacles to tight mazes with many narrow gaps. They found that this hybrid system was remarkably effective. By letting the fast intuition handle the easy or straightforward situations, the system avoided the heavy computational cost of the slow calculator for most of the time. In environments with large, open spaces or many small obstacles, the fast system solved the problems on its own more than half the time, saving significant processing power. When the fast system did get stuck or make a mistake, the slower system stepped in to fix it, ensuring the robot never crashed. The result was a robot that was almost as reliable as the slow, careful expert but much faster overall, successfully navigating complex paths in nearly every test case.
A key part of their discovery was how the robot learns over time. The researchers let the system run through hundreds of scenarios, and whenever the slow, careful system had to step in to fix a mistake made by the fast system, it saved that successful correction. The fast system then studied these corrections and improved its own intuition for the next round. This process, known as continual learning, allowed the fast system to get better and better, eventually solving more problems on its own without needing the slow system's help. However, the researchers also found a limit to this learning. The fast system improved most when the slow system provided short, local fixes, but it struggled to learn from the slow system when the solution required a long, complex plan that changed based on the entire route. This suggests that while the robot can learn to be faster, it still relies on the careful calculator for the most complicated, long-distance navigation tasks.
The team also investigated whether feeding the slow calculator the failed guess from the fast system would help it solve the problem faster. They found that this did not work as expected. Giving the slow calculator a bad starting point often just confused it or led it to the same dead end, offering no real time savings. This ruled out a common idea that simply passing along a partial solution always helps. Instead, the most efficient approach was to let the fast system try, verify the result, and only if it failed, let the slow system start fresh.
In the end, the study demonstrates that combining a fast, learned reaction with a slow, careful verification creates a robot that is both quick and safe. It does not require the fast system to be perfect on its own, nor does it force the robot to use the slow, expensive calculation for every single move. By carefully deciding when to trust a quick guess and when to demand a detailed plan, the system achieves a balance that neither method could reach alone. This approach offers a practical path forward for robots that need to move quickly and safely in the unpredictable, cluttered world of human life.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.