APQ Flagship Embodied Intelligence Controller TER40J-A6 in a Bipedal Humanoid Robot Application
Bipedal humanoid robots represent the culmination of embodied intelligence technologies. Their human-like form enables seamless adaptation to human working environments, giving them broad application prospects in industrial production, commercial services, and special operations. However, the dynamic balance challenges brought by the bipedal structure impose far higher requirements on the core control system than those of other robot forms—it must complete the full closed loop of perception, decision-making, and control within milliseconds.
After multiple rounds of evaluation and rigorous field validation, the customer ultimately selected APQ’s flagship embodied intelligence controller, the TER40J-A6, as the core control platform for its new-generation bipedal humanoid robot.
Customer Challenge
Extreme AI Computing Demand: The bipedal humanoid robot must run multimodal large models in real time at the edge to perform complex AI tasks such as environment perception, semantic understanding, and task planning, requiring AI performance at the thousand-TOPS level.
Synchronized Multi-Channel HD Video Acquisition: The robot needs multiple HD cameras to achieve 360-degree environmental perception, requiring the controller to support multiple GMSL2 camera interfaces and GNSS time-synchronization capability.
Coexistence of High-Speed Networking and Real-Time Industrial Signals: It requires 10GbE-class networking for large-scale data backhaul and training-data extraction, while also supporting the EtherCAT industrial real-time Ethernet protocol to drive all joint motors of the robot.
Real-Time Integrated Big-Brain/Small-Brain Control: Bipedal walking requires extremely short motion-control cycles. The communication latency of traditional “separated big-brain and small-brain” architectures cannot meet the real-time requirements of highly dynamic bipedal locomotion, making an integrated computing-and-control architecture essential.
Rich Industrial Communication Interfaces: The controller must simultaneously interface with joint servo motors, force sensors, IMUs, LiDARs, and other devices, requiring multiple CAN FD channels, serial ports, GPIO, and other extensive interfaces.
Vibration Resistance for High-Dynamic Operation: During highly dynamic movements such as running, jumping, and dancing, a bipedal humanoid robot is subjected to intense vibration and impact, requiring the controller to feature outstanding anti-vibration design and locking interface structures.
APQ Solution
2070 TFLOPS of AI Performance, Integrating Big-Brain and Small-Brain Functions:The TER40J-A6 is equipped with the NVIDIA Jetson Thor T5000 core module, delivering AI performance of up to 2070 TFLOPS (FP4 sparse). Compared with the Jetson AGX Orin platform, it represents a qualitative leap—with more CPU cores and approximately two to three times the CPU performance. This level of computing power enables bipedal humanoid robots to run large-scale vision-language models, multimodal perception models, and complex task-planning algorithms in real time at the edge, truly enabling end-to-end intelligent processing from “perception” to “decision-making."
8 GMSL2 Camera Interfaces for Panoramic Perception:The TER40J-A6 comes standard with 8 GMSL2 camera interfaces and supports GNSS time synchronization. These eight camera interfaces can directly connect to multiple HD cameras, enabling 360-degree blind-spot-free environmental perception for bipedal humanoid robots. GNSS time synchronization ensures precise alignment across multiple image streams, providing high-quality data support for core functions such as SLAM mapping, target tracking, and dynamic obstacle avoidance.
Unified Computing-and-Control Architecture, Eliminating Big-Brain/Small-Brain Separation:Built on the Jetson Thor platform, the TER40J-A6 is positioned as a flagship integrated big-brain/small-brain controller for embodied intelligence. It not only performs the “big-brain” role of AI perception and decision-making, but also undertakes the “small-brain” role of real-time motion control through its rich industrial communication interfaces, truly realizing a unified big-brain/small-brain design. This architecture fundamentally eliminates control lag caused by communication delay in traditional “big brain + small brain” separated solutions and provides a solid hardware foundation for highly dynamic real-time control in bipedal locomotion.
Comprehensive Industrial Communication Upgrade for Multi-Device Integration:The TER40J-A6 has undergone a comprehensive interface upgrade for bipedal humanoid robot applications:2× 10GbE ports + 4× independent Gigabit Ethernet ports (some of the Gigabit ports support EtherCAT): The 10GbE ports enable rapid extraction of large-model training data, providing high-quality material for continuous AI algorithm iteration and optimization. The Gigabit ports can be dedicated to joint-servo communication, LiDAR access, and backend communication, enabling traffic separation without congestion;4× CAN FD channels: Directly interface with the joint servo drive system of the bipedal humanoid robot to achieve multi-axis high-precision synchronized control;2× RS232/RS485 ports: Stably connect peripherals such as force sensors and IMUs;16× GPIO channels: Facilitate expansion for alarm outputs, status indications, and other actuator interfaces;Additional I2C, SPI, TTL, PPS_IN, and 2× PWM interfaces to meet in-depth customization requirements;3× USB 10Gbps + 1× USB 5Gbps ports for connecting high-speed storage, industrial cameras, voice modules, radar, force sensors, and other devices;1× HDMI display interface for local debugging and visualization.
Full-Interface Anti-Vibration Design for High-Dynamic Operation:During highly dynamic actions such as running, jumping, and dancing, the controller of a bipedal humanoid robot is subjected to continuous high-intensity vibration and impact. The TER40J-A6 features a full-chain, systematic anti-vibration design at the interface level:GMSL camera interfaces feature built-in latch-locking structures to ensure secure connections for multiple HD cameras during intense motion;Network interfaces adopt the Mini I/O form factor and also feature locking mechanisms, offering superior vibration resistance compared with traditional RJ45 connectors;USB interfaces provide reserved screw-mount positions around the ports, allowing cables to be further reinforced by fastening;This complete interface scheme ensures that the controller can stably support prototype robots in performing complex, high-dynamic, high-impact actions such as dancing and somersaulting.
Wireless Expansion and Industrial-Grade Reliability:The TER40J-A6 supports an M.2 Key-M slot for NVMe SSD expansion, an M.2 Key-B slot for 4G/5G module expansion, and an M.2 Key-E slot for Wi-Fi/Bluetooth module expansion. Reserved debugging and flashing interfaces, as well as a SIM card slot, are provided on the side for convenient maintenance and upgrading. The entire unit measures only 168.2 × 130 × 72.6 mm, uses intelligent active fan cooling, supports 12–28 VDC wide-range power input, offers an operating temperature range of -20 to 50°C and a storage temperature range of -40 to 80°C, and has passed rigorous vibration and shock testing.
Value Achievements:
Full Breakthrough of the Computing Bottleneck: With 2070 TFLOPS of ultra-high AI performance, multimodal large models can run in real time at the edge. Response speed for complex AI tasks such as environment perception, semantic understanding, and task planning has achieved a qualitative leap over traditional solutions, providing a powerful computing foundation for bipedal humanoid robots to evolve from merely “able to walk” to “able to think and make decisions.”
Unified Big-Brain and Small-Brain, Near-Zero Control Latency: The integrated computing-and-control architecture of the Jetson Thor platform completely removes the communication latency of traditional “big brain + small brain” separated solutions. Motion-control commands for bipedal walking can complete the loop from perception to execution within nanoseconds, greatly improving walking stability and dynamic response speed.
Blind-Spot-Free Panoramic Perception: The 8 GMSL2 camera interfaces, combined with GNSS time synchronization, enable precise acquisition and temporal alignment of multiple HD image streams, allowing the robot’s 360-degree perception capability to cover diverse scenario needs such as complex terrain recognition, dynamic obstacle avoidance, and target tracking.
Both High-Speed Communication and Real-Time Control: The 2× 10GbE ports ensure high-speed aggregation of HD video and perception data, enabling fast extraction of training data for large models and supporting continuous AI algorithm optimization. The 4× GbE ports support EtherCAT, enabling microsecond-level real-time control of joint motors.
Full-Interface Anti-Vibration Protection for High-Dynamic Tasks: Multiple anti-vibration measures—including latch-locked GMSL connectors, locking Mini I/O network ports, and screw-fastening positions around USB interfaces—ensure stable controller operation during highly dynamic actions such as running, jumping, and dancing, fundamentally solving persistent problems in traditional solutions where vibration causes loose interfaces and momentary signal interruptions.
Multi-Device Integration on a Single Platform, Greatly Simplifying System Architecture: Rich interfaces including 4× CAN FD, 2× RS232/485, 16× GPIO, and I2C/SPI/PPS enable a single-platform connection for joint servos, force sensors, IMUs, LiDARs, HD cameras, and other devices. Full system deployment can be completed without extra expansion modules, significantly reducing system integration complexity and BOM cost.
