Robots running in the wild? Can the new algorithm really adapt to an unpredictable environment

If a four-legged robot walking like a newborn foal, without stumbling is making you stand in awe, here is the news. They are up for an upgrade. The robots are running in the wild, quite literally! A research team from California developed a new robotic algorithm to give these legged robots a push. They can now walk and run and move around avoiding both stationary and moving objects. They have shared the code on GitHub for Robotics enthusiasts. Locomotive robotics, as a discipline has not only promoted stable, faster, and more efficient robotic movement but also enable a variety of walking and running behaviors. To make robotics practically viable, it is important that the tasks are optimized and the inability to integrate the sensory perceptions a robot depends on, has been a major stumbling block. While the current approach to training robots depends on proprioception or vision ie., it relies on either external sensing systems like lidar, cameras, and eyeballs or internal sensing that involves touch and force sensing, but cannot synchronize both. The new robotic algorithms, the researchers say train robots to identify asynchronous inputs from both visions to match them reasonably to decide quickly by anticipating changes in their environment ahead of time. Random delays ranging from 0.04 to 0.12 seconds were introduced to test the robots to evaluate their performance.

At the testing stage, they could achieve robots bypassing random obstacles such as sandy surfaces, gravel, grass, and bumpy hills covered with litter, poles, benches, moving boxes, and people, independently and swiftly. It could also navigate the office space bypassing boxes, desks, and chairs. The research paper will be presented at the Kyoto Conference on Intelligent Robots and Systems (IROS), which will be held between Oct 23 to 27 of 2022 in Kyoto, Japan. Experts believe, the algorithm will help engineers design robots for deploying for rescue operations, that need gathering critical information from dangerous terrains.

The algorithmic system the senior researcher Xiaolong Wang and his team developed can combine data gathered from real-time images taken by the camera mounted on the robot's head with sensor data from its legs. Having tested the moves in similar environments with random delays, it was found that the walking robot learned the moves, committing fewer mistakes, giving hope for better adaptation to the real world. The robots could identify objects moving in random directions at random speeds just because they have acquired the ability to understand the environment better.

Wang says, "The problem is that during real-world operation, there is sometimes a slight delay in receiving images from the camera, so the data from the two different sensing modalities do not always arrive at the same time." They call it Multi Modal Delay Randomization (MMDR), wherein the inputs from proprioceptive and visual states are intentionally placed out of order to train them on Reinforcement Learning policy. Further explaining Wang says, "In one case, it's like training a blind robot to walk by just touching and feeling the ground. And in the other, the robot plans its leg movements based on sight alone. It is not learning two things at the same time."

In the concluding remarks, highlighting latency as a critical gap, they say, the MMDR technique fairly addresses this issue for vision-guided quadruped locomotion control. And, both in simulation and on a real bot, the technique has proven to give reasonable outcomes, as robots have shown generalization and adaptation abilities towards unpredictable situations. Sounding optimistic about taking the technique further, the researchers say, "Right now, we can train a robot to do simple motions like walking, running, and avoiding obstacles. Our next goals are to enable a robot to walk up and down stairs, walk on stones, change directions and jump over obstacles." The developments, they opine, are not only an achievement for quadruped robotics but any branch of robotics that depends on visual robotic control.