Compile | Chen Junda
Edit | Panken
Zhidong reported on August 1 that on July 15, Disney released a 16-page paper detailing the main design and control methods of their new bipedal robot BD-X. The robot resembles "WALL-E" from Robot Story, and is inspired by the robot of the same name from "Star Wars". Its hardware uses standard parts and a 3D printed shell, and the legs have 5 degrees of freedom, and through reinforcement learning, it has mastered the walking posture of animated characters, and the movement is dexterous and smooth, and it has also exercised the ability to cross complex terrain.
At the IEEE event not long ago, this Disney robot also showed off its skills face-to-face with the robot dog of Hangzhou Unitree Technology.
▲Disney BD-X robot interacts with Unitree Go2 (Source: YouTube)
Earlier, it appeared as a surprise guest at the NVIDIA GTC conference in March this year, interacting intimately with NVIDIA founder and CEO Jensen Huang, and shaking his head and selling cuteness.
▲The BD-X robot, which resembles WALL-E, interacts with Jensen Huang at the GTC conference (Source: NVIDIA)
The robot's entertainment design concept greatly reduces hardware complexity and cost. But Disney has added additional speakers, antennas, headlamps and eyes to the robot, which allows it to express its "emotions" in more ways.
While BD-X seems to be alive and can really understand what people are saying, it doesn't have the ability to actively perceive and behave. Its every move is achieved by human operation. The robot is equipped with Jetson, an on-board computer from NVIDIA, which combines human instructions with the experience gained during training to output final action instructions.
▲Screenshot of the Home of the paper (Source: Disney)
Address:https://la.disneyresearch.com/publication/design-and-control-of-a-bipedal-robotic-character/
1. The mechanical design is simple and effective, serving entertainment and creative purposes
Disney chose a different path when designing the BD-X robot than most businesses. Rather than seeking the best and most powerful mechanical design, they opted for simple designs that would serve both creative and entertaining purposes. The researchers successfully integrated elements such as animation content, design, control, and real-time manipulation, and achieved the rapid development of custom robot characters in less than one year.
The robot mainly uses standardized parts on the market, while the shell is 3D printed, but after software debugging, it can still achieve good performance while reducing hardware complexity.
▲Disney BD-X robot design drawing (source: Disney)
The bipedal robot has a total weight of 15.4 kg and a height of 0.66 meters. There are 5 degrees of freedom on each leg and 4 degrees of freedom for the head and neck. The high degree of freedom of the legs allows the robot to perform a wide range of movements, while the high degree of freedom of the head and neck allows the robot to perform a variety of interesting and emotionally rich movements through movements.
▲The BD-X robot was unveiled at Disneyland, and the two people on the far right of the screen are operators (Source: YouTube)
Disney's engineers also cleverly designed the robot to achieve a realistic knee bend effect without a knee bending actuator and relying on contact with the ground. The robot is equipped with a microcontroller-driven communication board that communicates with the built-in computer, actuator and inertial measurement unit at a frequency of 600 Hz. 1 removable battery powers the robot for at least 1 hour of operation.
Interestingly, in addition to the speakers on the head and torso, the robot also has antennas with actuators, eyes that light up, and a headlamp. These devices provide more ways for the robot to express their emotions, and their control and drive are relatively independent, so you can choose whether to carry these functions or not.
2. Use reinforcement learning to master animation movements and simulate training in NVIDIA Isaac Gym
To provide a clear framework for robot movements, Disney researchers divided robot movements into the following three categories.
1. Sustained action: There is no clear beginning and end point. The robot balances and responds to the measured state and a continuous flow of control inputs.
2. Periodic Motion: There is a periodic phase signal that is passed to the strategy. In this mode, the phase signal loops indefinitely.
3. Occasional exercise: there is a predetermined duration. The strategy receives a monotonically increasing phase signal, and as soon as the movement is over, it forces a transition to a new one.
▲3 types of movement of BD-X robots (Source: Disney)
The control strategy of a robot is an important concept in robotics, which refers to the mapping of the state of the robot as a function of actions or motor commands, so that the robot can achieve the expected behavior and tasks.
The researchers allowed the BD-X robot to master a variety of control strategies through reinforcement learning, targeting the above three types of actions.
Based on the robot's CAD model, Disney researchers constructed the robot's accurate physical form, actuators, and interactions with the environment, and simulated the robot's rigid body dynamics using NVIDIA's robot training platform, Isaac Gym. The researchers further simulated the full dynamics of the robot through a customized actuator model.
▲BD-X simulation training system (Source: Disney)
The movements of the BD-X robot are learned based on the content of the animation. Disney researchers extracted information about the animation's character-specific movements, such as the global position of the torso, the position of the joints, and vectors, which were used to train the BD-X robot to mimic the animation's movements.
In Isaac Gym's simulation training, the robot receives rewards and punishments from the reinforcement learning system. The system takes into account the similarity of the movements to the animated characters, the degree of regularity of the movements (e.g., whether the joints are twisted, whether the movements are fluid), and whether the robot is alive ("death" is defined as the head and torso touching the ground, or the head and torso colliding), and gives corresponding rewards and punishments.
During training, the system also simulates random disturbances, traction, and different terrain formations, which improves the robot's robustness (meaning anti-interference or stability).
3. There is no active perception and action ability, and movement is realized through complex controllers
Disney researchers said that at the time of publishing the paper, the BD-X robot had accumulated 10 hours of public display runtime, during which the robot performed perfectly and did not fall once. Many viewers were attracted by the robot's cute appearance and did not notice the robot's manipulator.
In fact, the BD-X robot does not have any active behavior and perception capabilities, and its actions are realized by a complex controller operated by a professional. During operation, the operator can control the robot using an intuitive remote-controlled interactive interface. This control device is more complex, with 2 joysticks, 2 path boards, and a total of more than a dozen buttons on the front and back.
▲BD-X realizes flexible movements through the controller (Source: Disney)
At the end of the virtual training, the weights of the neural control strategy are frozen, and the strategy network is deployed to the robot's onboard computer, Jetson. Once deployed, the control strategy and low-level controller interact directly with the robot's hardware and can also incorporate measurement data from inertial measurement units and actuators.
The robot's built-in animation engine automatically combines the operator's input with the robot's built-in strategy control commands, display function signals (antenna, eye, headlamp control signals) and sound signals to generate final instructions for the control strategy.
▲The operator debugs the BD-X robot in the background of NVIDIA GTC 2024 (Source: YouTube)
The combination of manual manipulation instructions and built-in policy control commands can avoid the actions that researchers want to avoid during training, such as falling, stiff and unsmooth movements, and so on.
▲The BD-X robot shows strong robustness and can remain stable in a variety of complex ground situations (Source: Disney)
Presentation features and audio are controlled elements of the bot that play a key role in representing the character without affecting the dynamics of the system. Their behavior is synchronized with the robot's movements through animation signals and state feedback from the animation engine.
Conclusion: AI brings animated characters into reality, and "impractical" robots are also valuable
In fact, Disney is a veteran player in the field of robotics, and they have been deeply involved in robot-related research since the 60s of the last century, and they have tried to deploy their research results in Disneyland to explore various ways and scenarios that bring animated characters into the real world.
▲Disney's past research in the field of robotics (Source: Disney)
The development of AI technology has made every move of robots more realistic and natural, while reinforcement learning has allowed the actions of characters made by professional animators to be implemented in reality. Although BD-X robots are not as capable of entering factories and assembly lines as other bipedal robots, we are also happy to see the further development of robotics technology for entertainment purposes, so that ordinary people can also enjoy the happiness brought by robots.
Source: Disney