Toyota Research Institute has a new product! A soft humanoid robot is here!

In the past year, humanoid robots have become the focus of attention, and many companies have lined up to launch their own humanoid products. Most of them have a typical humanoid appearance, use arms and grippers to handle objects, and hard legs are their way of walking.

But recently, Japan's Toyota Research Institute (TRI) launched a new robot Punyo, and expressed the hope that Punyo will take humanoid robots one step further.

Punyo is innovative in both the design concept and the way it is operated. It has no legs, and so far the TRI team has only been working on the robot's torso and developing manipulation skills.

Design concept: serving human daily life

Traditional industrial robots are mostly used in workshop operations, assembly and other tasks, with the aim of improving production efficiency and reducing labor intensity. In the future, service robots may enter more homes, directly facing and serving the daily needs of ordinary people.

TRI researchers said Punyo aims to be a robot that "helps people complete everyday tasks at home and elsewhere."

This design concept determines that Punyo needs to be flexible, soft, and safe. Because it needs to enter the complex and ever-changing home environment, it cannot have hard, rigid mechanical arms like traditional industrial robots, otherwise it will give people a sense of danger and will not be able to complete various daily object operation tasks. This is somewhat similar to the design concept of SoftBank's robot Pepper, both of which focus on how to make robots more integrated into human life.

Service-oriented applications also require Punyo to be able to learn a variety of daily skills, rather than just performing a single operation on a factory assembly line. This requires giving the robot a strong learning ability to master the operation methods of various daily tasks by observing and imitating human demonstrations.

Manipulation using the entire body is tricky for humanoid robots because balance is a difficult problem. However, TRI researchers designed its robot to do just that.

“Punyo does things differently. Using its entire body, it can carry more than simply pressing with an outstretched hand,” added Andrew Beaulieu, one of TRI’s full-body manipulation technology leaders. “The softness, tactile sensing, and ability to make a lot of contact allow for better manipulation of objects.”

Soft and hard combination body

To achieve a flexible and soft robot design, TRI used a combination of hard and soft robotic arm design. Punyo's hands, arms and chest are covered with compliant materials and tactile sensors , which can sense external contact, and the soft materials also allow the robot's body to conform to the object it manipulates.

This is a typical design idea for many current soft robots.

At the same time, under the soft shell, Punyo also retains two "hard" mechanical arms as skeletal support, as well as a torso frame and waist actuators to provide mechanical support and precise control. This hard-soft combination design combines the mechanical advantages of traditional robots with the soft characteristics of soft robots.

Specifically, the airbags on Punyo's arms can adjust the internal pressure to become hard or soft as needed, while ensuring a certain mechanical rigidity and providing about 5 cm of compliance. The "claw" also uses a high-friction latex airbag design, and the camera in the palm of the hand can sense the size of the external force by observing the deformation of the airbag surface. The entire arm can be bent and rotated, and the airbags are connected to each other, which allows the force to be transmitted smoothly to prevent the robot from "breaking its arm".

Strong learning ability

To adapt to the changing tasks in the home environment, Punyo must have strong learning abilities.

According to the TRI team, Punyo learned contact-rich policies using two approaches: diffusion policy and example-guided reinforcement learning. TRI announced its diffusion policy approach last year. With this approach, the robot uses human demonstrations to learn robust sensorimotor policies for difficult-to-model tasks.

Example-guided reinforcement learning is an approach that requires modeling a task in simulation and guiding the robot’s exploration with a small set of demonstrations. TRI said it uses this type of learning to achieve robust manipulation policies for tasks that can be modeled in simulation.

When a robot can see demonstrations of these tasks, it can learn them more effectively. It also gives the TRI team more room to influence the movement style the robot uses to complete tasks.

The team used Adversarial Motion Priors (AMP), traditionally used to stylize computer-animated characters, to incorporate human motion imitation into their reinforcement pipeline.

Reinforcement learning does require the team to model the task in simulation for training. To do this, TRI used a model-based planner for the demonstration, rather than teleoperation. It calls this process "plan-guided reinforcement learning."

TRI claims that using the planner can make long-distance missions that are difficult to operate remotely possible. The team can also automatically generate any number of demonstrations, reducing its pipeline’s reliance on human input, which brings TRI closer to increasing the number of missions that Punyo can handle.

Although Punyo software service robot is still in its early stages and its performance in all aspects needs to be improved, its application prospects are broad, and Punyo's design concept and technical route also provide new ideas for the industry.