R1-A5-D Ultimate B — mobile-base humanoid with Dex3-1 three-finger hands, wrist RGB cameras, MID360 LiDAR & stereo vision

What's in This Configuration

The R1-A5-D Ultimate B upgrades from the entry-level Ultimate A with Dex3-1 three-finger dexterous hands and wrist-mounted RGB cameras — while retaining the same NVIDIA Jetson Orin NX 100 TOPS compute module with 16 GB LPDDR5 and dual DLA cores. The Dex3-1 hands add an independent opposing thumb to the two-finger architecture, enabling tripod grasps, in-hand prehension, and dexterous manipulation of irregular objects while navigating. Wrist cameras provide close-up visual feedback for grasp verification and visual servoing. This is the configuration for researchers who need dexterity beyond parallel grippers with compute headroom for advanced perception and learning.

Dex3-1 — Three-Finger Dexterity

The Dex3-1 is a three-finger dexterous hand that adds an independent opposing thumb to the two-finger architecture. This third degree of freedom unlocks tripod grasps for spherical objects, in-hand prehension for reorienting items without setting them down, and dexterous manipulation of irregular shapes that defy parallel-jaw grasping. Each finger is independently actuated with position and velocity control, enabling adaptive force distribution across three contact points. On a mobile platform, this means the robot can pick up a ball, rotate it to align a connector, and insert it — all while navigating between workstations. With the Orin NX 100 TOPS, the Dex3-1 hands can run learned grasp policies, visual servoing, and tactile feedback loops that would be impossible on lesser compute.

Dex3-1 three-finger dexterous hands with wrist RGB cameras on the R1-A5-D mobile base

Wrist-Mounted RGB Cameras — See From the Gripper

Ultimate B adds dedicated RGB cameras to each wrist — providing a close-up view of the grasp target that the head-mounted binocular camera cannot match. These wrist cameras enable visual servoing for precise alignment during approach, grasp verification to confirm a stable hold before lifting, and in-hand tracking to monitor object pose during transport. Combined with the head stereo camera and MID360 LiDAR, the robot builds a multi-scale perception stack: LiDAR for global navigation, head camera for scene understanding, and wrist cameras for manipulation detail. The Orin NX 100 TOPS processes all three streams concurrently, with DLA cores handling vision networks while the GPU fuses multi-camera data for closed-loop control.

Vision-LiDAR-Wrist Fusion — Three-Scale Perception

Ultimate B processes synchronized streams from three sensor modalities on the Orin NX 100 TOPS: close-up RGB from wrist cameras, scene-level depth from the head binocular camera, and 360° point clouds from the MID360 LiDAR. The dual DLA cores run wrist-camera object detection and head-camera segmentation in parallel while the GPU handles LiDAR SLAM and multi-camera fusion — no resource contention even under full load. With 16 GB RAM, the system can hold multiple perception models in memory: a wrist-camera grasp detector, a head-camera scene parser, a LiDAR localization network, and a multi-modal fusion policy — all running concurrently. This three-scale perception enables research in visual mobile manipulation, active perception, dexterous grasping on the move, and embodied AI with rich sensory feedback.

Autonomous Mobile Base — Navigate, Then Manipulate

The R1-A5-D chassis adds 3 degrees of freedom to the upper body: a telescopic pillar for height adjustment and a 2-DOF wheeled base for omnidirectional navigation. Combined with the MID360 LiDAR and stereo camera head, the robot builds real-time maps, plans paths, and avoids obstacles autonomously — then stops precisely to manipulate. The removable lithium-ion battery in the chassis delivers approximately 1.5 hours of untethered operation.

Orin NX 100 TOPS — Flagship Compute for Entry-Level Manipulation

The built-in NVIDIA Jetson Orin NX delivers 100 TOPS of AI compute across an 8-core CPU, 16 GB LPDDR5, and dual NVDLA V2.0 accelerators. With three-finger dexterous hands and wrist cameras, this compute headroom enables advanced visual mobile manipulation: run real-time SLAM with MID360 LiDAR, process stereo depth and wrist-camera streams simultaneously, deploy learned grasp policies for tripod and precision grasps, and experiment with active perception and multi-robot coordination — all at once. The dedicated DLA cores offload vision networks from the GPU, leaving GPU resources free for sensor fusion, grasp planning, and dexterous control. The 16 GB RAM provides space for large models that would not fit on an 8 GB module. No external GPU server, no cloud dependency — just power on and deploy.

Binocular Vision & Scene Perception

A binocular camera head delivers a 146° × 124° field of view with a 60 mm stereo baseline for scene-level understanding. RGB streams at 1280×720@30 Hz and depth at 544×448@10 Hz. Combined with the MID360 LiDAR and wrist-mounted RGB cameras on each Dex3-1 hand, the R1-A5-D Ultimate B offers a three-scale perception stack for visual mobile manipulation research — with the Orin NX 100 TOPS processing all streams concurrently via DLA-accelerated vision networks.

Voice Collaboration & Natural Interaction

Four-array beamforming microphones and dual 3 W speakers enable clear voice capture and synthesis in noisy labs, warehouses, or classrooms. The open audio framework supports custom wake words, multilingual TTS, and integration with LLMs for conversational command-and-control — even while the base is in motion.

Collapsible. Mobile. Ready to Deploy.

At 683 × 520 × 440 mm collapsed (1323 mm elevated), the R1-A5-D Ultimate B navigates standard doorways, elevators, and lab aisles. The telescopic pillar lowers the upper body for transport and raises it for manipulation. At approximately 30 kg with battery, it is a self-contained mobile research platform — no external compute rack, no tether cables during operation.

Collapsed (683 mm) and elevated (1323 mm) configurations for navigation and manipulation

Full-Stack Open Development

Unitree provides a mature, Linux-based robot development framework with open APIs for the underlying system, robotic arms, mobile base, audio, lighting, and visual control. ROS 2 native support, dual encoders on every arm joint, and hollow internal wiring make the R1-A5-D Ultimate B a clean slate for reinforcement learning, dexterous mobile manipulation research, embodied AI, and active perception studies. The built-in Orin NX 100 TOPS with dual DLA cores and 16 GB RAM handles multi-camera perception, grasp planning, and dexterous control locally — with compute headroom to grow into tactile sensing, larger models, and multi-robot coordination without hardware limits.

Upgrade Path — Grow With Your Research

Ultimate B is the entry point with flagship compute already installed. The Orin NX 100 TOPS, 16 GB RAM, and dual DLA cores provide headroom to grow:

• End Effectors — Upgrade to Dex3-1, BrainCo Revo 2, LinkerBot O6, or tendon-driven hands as your manipulation needs evolve. The compute is already ready.
• Wrist Cameras — Add RGB cameras to the grippers for close-up visual feedback and grasp verification.
• Tactile Sensing — Integrate force-sensitive fingertips for contact-aware grasping.
• Large Vision-Language Models — Run VLA policies and multi-modal foundation models locally without quantization.

Designed For

Technical Specifications

What's in the Box

• Unitree R1-A5-D Ultimate B Humanoid Robot (mobile base)
• Dex3-1 3-finger dexterous hands with wrist RGB cameras (pre-installed, both arms)
• NVIDIA Jetson Orin NX 100 TOPS compute module with 16 GB RAM & dual DLA cores (pre-installed)
• Removable chassis lithium-ion battery pack
• External power supply & AC cable
• MID360 LiDAR module (pre-installed in chassis)
• Binocular vision head module (pre-installed)
• 4-array microphone + dual speaker module (pre-installed)
• Standard dual-hand remote controllers
• Quick-start guide & safety manual
• Secondary development documentation access

R1-A5-D Configuration Guide

Every configuration shares the same R1-A5-D mobile base — the difference is the end effector and compute module. Choose the one that matches your research goals: