Simulating humans into robots in nature
Significant advances in robot manipulation recently achieved. However, most testing is done in a lab setup or simulation, not in a real-world environment.
A recent article on arXiv.org proposes visual imitation of humans as a safe and scalable alternative to general robotic manipulations in the wild.
First, advances in computer vision and computer photography are used to infer trajectories and interactive information from humans. Then, a policy optimization strategy based on sampling and agentless representations is used to improve the policy iteratively through real-world interactions. The unknown task discovery policy allows delegates to sample new and interesting actions.
A demonstration showing success in various real-world tasks such as opening and closing doors, folding a shirt, or wiping a whiteboard.
We approach the problem of learning by observing people in nature. While traditional approaches in Imitation and Reinforcement Learning hold promise for real-world learning, they are either model ineffective or limited in laboratory settings. Meanwhile, there has been a lot of success in dealing with passive, unstructured human data. We propose to solve this problem through an efficient one-time robotic learning algorithm that focuses on learning from a third-person perspective. We call our method WHIRL: Learning robots to mimic humans in nature. WHIRL pre-extracts protester intent, using it to initialize our agency’s policy. We introduce an effective real-world policy learning plan that improves the use of interactions. Our key contributions are a simple sampling-based policy optimization method, a new objective function for aligning human and robot video, and a exploratory method to increase sampling efficiency. sample. We demonstrate generality and success in one go in real-world settings, covering 20 different manipulative tasks in the wild. Video and talk at This https URL
Research articles: Bahl, S., Gupta, A., and Pathak, D., “Mimicking Humans into Robots in Nature”, 2022. Links: https://arxiv.org/abs/2207.09450