Disney teaches robots how to fall properly (VIDEO)

Two-legged robots are complex and expensive machines, but when they fall, they usually don't behave gracefully at all: one awkward movement or jolt and the robot crashes hard to the floor, risking damage to cameras and other sensitive components.

A team at Disney Research in Zurich, Switzerland, has introduced a system that teaches a robot to fall gently, controllably and even beautifully, Techxplore reports.

Traditional approaches to protecting robots when they fall are ineffective. Actuators can "freeze," turning the robot into a rigid structure that hits the floor with maximum force. The alternative is to completely immobilise the mechanics, allowing the robot to tumble as it pleases, but then the motion becomes a chaotic fall. Another option is pre-programmed fall scenarios that work only in simple, slow movements and predictable situations.

In a new paper posted on the arXiv preprint server, Disney engineers describe how they used reinforcement learning to teach a robot how to fall safely and "stylised". They ran thousands of virtual bipedal robots in a computer simulator and had them fall from a variety of angles and poses.

For each fall, the virtual robot was 'rewarded' for two key outcomes: minimising the force of impact on the surface and achieving a given final pose from a predefined set. The constant feedback allowed the system to develop a general strategy for 'crashing gracefully' - regardless of exactly how the fall started.

The resulting control policy was then transferred to a real bipedal robot. The team chose 10 spectacular final poses that the artists had designed in advance. The robot was deliberately pushed or prodded with a stick from different sides and at random angles to provoke a fall.

As a result, the robot fell repeatedly, but suffered no noticeable damage and remained fully functional. Each time, it performed a controlled movement, trying to complete the fall in one of the given artistic poses.

"Our method provides soft fall behaviour and protects the most sensitive parts of the robot regardless of the direction of the fall," the authors write. This is the first general approach to demonstrate user-controlled falls of a bipedal robot in a real-world setting, they said.

The Disney team now plans to test how universal the approach is. The plan is to test it on other types of robots, such as four-legged robots, to see if they can be taught safe falls in the same way.

Another challenge is to teach the robot to predict the onset of a fall. At present, the system is only switched on after a fall has already started. The researchers want the algorithm to recognise the loss of balance in advance and launch a protective manoeuvre before the critical moment.

In addition, the developers intend to teach the robot not only to fall beautifully, but also to rise from the floor as confidently and smoothly - that is, to close the cycle of "falling and getting up", preserving the impression of a lively, natural movement.