Simulated humanoids provide an interesting platform for investigating motor intelligence with the ability to mimic the full spectrum of human motion. An important research area of machine learning is the acquisition and application of motor skills. Physically simulating human talent presents significant control challenges. Controllers must deal with high-dimensional, unstable, and discontinuous systems that require precise timing and coordination to achieve the desired motion.
All current learning methods make it difficult to master complex humanoid behaviors using the Tabula Rasa approach. Motion capture (MoCap) data has rapidly become an integral part of humanoid control research. A MoCap trajectory is a sequence of configurations and poses assumed by the human body throughout the motion in question and contains kinematic information about the motion. This information helps simulated humanoids learn basic motor skills through motion capture demonstrations, facilitating training of complex control strategies.
Unfortunately, using motion capture data in a physics simulator requires recovering the actions (such as joint torques) that produce a series of kinematic poses on a given motion capture trajectory (trajectory tracking). Finding an action sequence that turns a humanoid into a motion capture sequence is not easy. Reinforcement learning and adversarial learning are two approaches that have been used to address this problem. Training an agent to recreate hours of motion-capture data also requires a large amount of computing, and the computational load for detecting these actions scales with the amount of motion-capture data. increase. Although MoCap datasets are widely available, few research organizations have significant computational resources that can use them to further advance learning-based humanoid control.
A recent Microsoft study introduced MoCapAct, a dataset of high-quality MoCaptracking rules for MuJoCo-based simulated humanoids, and a collection of rollouts from these expert policies.
MoCap is designed to be compatible with the very popular dm_control humanoid simulation environment, with the aim of removing the current barriers and making MoCap data usable in humanoid control research. CMU MoCap is one of the largest publicly available MoCap datasets. MoCapAct’s policy can track all his 3.5 hours of that data.
The researchers explored the expert policies of MoCapAct and used expert rollouts to train a single hierarchical policy that could track all motion capture clips considered, thereby providing a method for learning different motions. Demonstrate the use of MoCapAct. Low-level parts of the policy are reused for efficient RL task learning.
The team trained a GPT network to generate motions in the MuJoCo simulator in response to motion prompts using a dataset for generative motion imputation.
This dataset allows the research group to avoid the time- and energy-consuming process of using MoCap data to learn low-level motor skills. This greatly reduces the entry hurdles for simulated humanoid control and promises a wealth of possibilities for exploring multitask learning and motor intelligence. The team believes this approach can be used in training alternative policy frameworks such as Decision Transformer, and in setups such as offline reinforcement learning.
This Article is written as a research summary article by Marktechpost Staff based on the research paper 'MoCapAct: A Multi-Task Dataset for
Simulated Humanoid Control'. All Credit For This Research Goes To Researchers on This Project. Check out the paper, github, project page and reference article.
Please Don't Forget To Join Our ML Subreddit
” data-medium-file=”https://www.marktechpost.com/wp-content/uploads/2020/10/Tanushree-Picture-225×300.jpeg” data-large-file=”https://www.marktechpost.com/wp-content/uploads/2020/10/Tanushree-Picture-768×1024.jpeg”/>
Tanushree Shenwai is a consulting intern at MarktechPost. She is currently pursuing her bachelor’s degree at the Indian Institute of Technology (IIT), Bhubaneswar. She is a data her science enthusiast and has a keen interest in the scope of artificial intelligence applications in various fields. Her passion lies in exploring new advancements in technology and its practical applications.
Comments
Post a Comment