DEEP Robotics is a leader in embodied AI technology innovation and application, being the first in China to achieve fully autonomous inspection of substations with quadruped robots.
DEEP Robotics’ self-developed robot series have been successfully deployed in various applications, including power station, factory, and tunnel inspections, as well as emergency rescue, fire detection, and scientific research. DEEP Robotics’ robots have served in underground tunnels for the Asian Games, participated in the Singapore National Grid project, and engaged in drill missions such as earthquake disaster relief drill “Emergency Mission” and explosion detection drills. Currently, DEEP Robotics is engaged in long-term, in-depth cooperation with industry giants such as State Grid Corporation of China, China Southern Power Grid, Baosteel, and Fluke.
Key Robots;
X30
The X30 is a flagship quadruped robot tailored for industrial applications such as inspection, security, and mapping. It features a maximum speed of 4 m/s, can navigate slopes up to 45°, and overcome obstacles up to 20 cm in height. With an IP67 ingress protection rating, it operates in temperatures ranging from -20°C to 55°C. The robot offers a battery life between 2.5 to 4 hours, covering a distance of up to 10 km. Communication interfaces include Ethernet and WiFi, with output power supply options of 5V, 12V, and 24V.
X30 Pro
An enhanced version of the X30, the X30 Pro includes additional communication interfaces such as USB 2.0 and USB 3.0, providing greater flexibility for various applications.
Lite3
The Lite3 is a compact and lightweight quadruped robot designed for agility and versatility. It features high-torque joint drive modules, a runtime of 1.5 to 2 hours, and a maximum speed of 2.2 m/s. The robot can navigate slopes up to 40° and stairs up to 15 cm in height.
Lynx
The Lynx is an all-terrain, off-road robot capable of climbing 22 cm steps and tackling 45-degree slopes. It features a wheeled-leg hybrid design, combining speed and agility, with a battery life of up to 3 hours. The robot is priced under $20,000, making it accessible for various applications.
Humanoid Robot
DEEP Robotics has unveiled its first humanoid robot, designed to assist humans in performing repetitive, high-precision tasks. Through AI and big data training, these robots are expected to safeguard human health and safety in various environments.
The core team members are from well-known universities, those are Zhejiang University, Shanghai Jiao Tong University, Beijing Institute of Technology, Wuhan University, University of Electronic Science and Technology of China, University of Chinese Academy of Sciences, New York University, University of Illinois at Urbana-Champaign and Georgia Institute of Technology.
Proximal Policy Optimization (PPO) method combined with depth camera sensor inputs and trained an end-to-end policy.
The Proximal Policy Optimization (PPO) algorithm, combined with depth camera sensor inputs, is often used in robotics and AI for solving tasks involving perception and control.
Proximal Policy Optimization (PPO) Overview
PPO is a state-of-the-art reinforcement learning (RL) algorithm developed by OpenAI, known for its stability and efficiency. It operates on the principle of improving policies through gradient-based optimization while ensuring that updates are constrained to prevent performance collapse.
Key Characteristics of PPO:
Policy Gradient Method: PPO optimizes policies directly using gradients, which allows it to learn complex behaviors.
Clipped Objective Function: Ensures updates stay within a safe range, balancing exploration and exploitation.
On-Policy Learning: Uses data collected from the current policy to optimize future behavior.
Efficiency: Simple to implement and computationally efficient compared to other RL algorithms like Trust Region Policy Optimization (TRPO).
Depth Camera Sensor Inputs
A depth camera sensor provides 3D perception by capturing depth information of the environment. It outputs depth maps or point clouds that can be processed to understand spatial relationships, distances, and object geometry.
Applications in Robotics:
Obstacle Avoidance: Detecting and navigating around obstacles. 3D Mapping: Constructing a 3D representation of the environment for path planning. Manipulation: Identifying objects in cluttered spaces for picking or interaction.
Localization: Enhancing localization through depth-based feature matching.
Combination of PPO with Depth Cameras Integrating PPO with depth camera inputs involves using the 3D spatial data as part of the state representation in reinforcement learning. The agent uses this information to make decisions and learn optimal policies for a given task.
Sensor Data Processing:
Depth maps from the camera are preprocessed to extract features (e.g., edges, regions of interest).
Techniques like convolutional neural networks (CNNs) or point cloud processing may be used.
State Representation:
The processed depth information is combined with other sensor inputs (if any) to form a state vector.
Policy Learning:
PPO is trained to map states to actions by interacting with the environment.
The reward function is designed based on the task (e.g., minimize collisions, maximize coverage).
Optimization:
The PPO algorithm optimizes the policy while ensuring updates do not diverge significantly, aided by the clipped objective.
Use Cases:
Autonomous Navigation:
Depth cameras enable real-time obstacle detection and path planning.
PPO learns policies for navigating complex environments.
Robot Arm Manipulation:
Used for tasks requiring fine motor control and precise object interaction.
Depth cameras provide 3D object localization and segmentation.
Inspection and Monitoring:
Robots equipped with depth cameras can inspect infrastructure or machinery.
PPO learns inspection paths optimized for coverage and time efficiency.
Dynamic Object Interaction:
For tasks like catching or pushing objects, PPO learns how to adapt movements dynamically using depth inputs.
Benefits:
Improved Perception: Depth cameras provide richer environmental understanding compared to traditional cameras.
Efficient Learning: PPO’s stability ensures faster convergence even in high-dimensional state spaces.
Generalization: PPO can generalize well across similar tasks when trained with diverse depth inputs.
This combination is widely used in robotics, autonomous systems, and gaming for tasks requiring both complex decision-making and rich sensory input.