The west wind comes from the Wafei Temple
量子位 | 公众号 QbitAI
Shooting, dribbling, finger spins... This physics-simulated humanoid robot will play ball:
There are many tricks to know:
After a show of skills, it turned out that they were all learned from others, and every action detail was accurately copied:
That's why a new study called PhysHOI has recently been able to allow physically simulated humanoid robots to Xi learn and mimic these movements and techniques by watching demonstrations of human-object interaction (HOI).
The point is that PhysHOI does not need to set a specific reward mechanism for each specific task, and the robot can learn Xi and adapt autonomously.
In addition, there are a total of 51x3 independent control points on the robot's body, so it can be highly realistic to imitate.
Let's take a look at how it works.
Simulated humanoid robot transforms into a "slam dunk"
The work was co-authored by researchers from Peking University, the IDEA Institute, Tsinghua University, and Carnegie Mellon University.
According to the researchers, most of the previous similar work had limitations such as the isolation of imitation movements, the reward that required specific tasks, and the whole body movement that did not involve dexterity.
The PhysHOI they proposed solves these problems by using motion capture technology to extract HOI data, and then using the imitation Xi to Xi learn human movement and object control.
Among them, one of the important components of HOI data is the kinematic data covering human movement, object motion, and relative motion, which records information such as position, velocity, and angle.
In addition, it is important to have dynamic data that reflects real-time changes and updates during movement.
In order to make up for the lack of dynamic information in HOI data, the researchers introduced contact graph (CG).
The CG node is composed of the robot's limb parts and objects, and each edge is a binary contact label that only expresses the two states of "contact" or "non-contact".
In addition, multiple limb parts can be placed in a single node to form an aggregated CG.
Specifically, the PhysHOI approach is:
Firstly, the reference HOI state sequence is obtained through motion capture, including human motion, object motion, interaction diagram and contact diagram.
Then, the information from the first frame is used to initialize the physical simulation environment and build a system state containing the current simulation state and the next reference state.
Next, the action generated by the strategy network is input to control the humanoid robot, and the physics simulator updates the state of the human body and objects in the scene according to the action, and calculates the rewards including motion matching, contact map and other aspects.
The strategy network is optimized by using reward, state and action samples, and the updated policy network is used to start a new round of simulation process, and so on until the network converges, and finally the control strategy that can reproduce the reference HOI skill is obtained.
It is worth mentioning that the researchers designed a task-independent HOI imitation reward, which does not need to customize the reward function for different tasks, including the body and object reward that reflects the motion matching degree, the contact graph reward that reflects the contact correctness, and avoids the local optimal solution such as using the wrong body part to touch the object.
Contact graph rewards are key
The researchers tested PhysHOI on two HOI datasets.
The BallPlay dataset is introduced, which contains a variety of full-body basketball skills.
The researchers selected 5 grabs cases in the S8 subset of the GRAB dataset and 8 basketball skills in the BallPlay dataset.
Using the previous methods such as DeepMimic and AMP as baselines, the researchers modified them to adapt to the HOI imitation task for fair comparison.
The results showed that the previous method of using only kinematic rewards could not accurately reproduce the interaction, and the ball would fall or fail to grasp.
Under the guidance of the contact map, PhysHOI successfully carried out HOI imitation.
PhysHOI achieved the highest success rates of 95.4% and 82.4% on both datasets, respectively, while also achieving the lowest motion error, which was significantly better than other methods.
Ablation studies show that the contact graph reward can effectively avoid falling into local optimality by using only motion information, and guide the robot to achieve correct contact.
Without the contact graph reward, the humanoid robot may not be able to control the ball, or mistakenly use other parts of its body to control the ball:
Paper link: https://arxiv.org/abs/2312.04393
— END —
QbitAI · Headline number signed
Follow us and be the first to know about cutting-edge technology trends