Unitree G1 Ultimate A Humanoid Robot with 3 Finger Hands (G1EDU-U3)

Unitree G1 Ultimate A Humanoid Robot with 3 Finger Hands (G1EDU-U3)

The G1 EDU Ultimate A (U3) includes full secondary development support through Python, C++, and ROS 2 SDK access, the NVIDIA Jetson Orin NX 16GB (100 TOPS) compute platform, LIVOX MID-360 3D LiDAR, Intel RealSense D435i depth camera, WiFi 6 and Bluetooth 5.2 connectivity, and the 9,000 mAh quick-release battery providing approximately two hours of runtime.

The U3 is the first configuration in this progression that includes actual object manipulation hardware. The U1 and U2 configurations provide locomotion, arm articulation, and SDK access without end-effectors capable of grasping. The U3's Dex3-1 hands enable the robot to pick up, transport, and manipulate physical objects — the fundamental capability that distinguishes manipulation research from locomotion and HRI research.

G1 Foldable Frame: 132 cm, 35 kg, Portable

The G1 EDU Ultimate A uses the standard G1 physical frame. Standing dimensions are 1320 x 450 x 200 mm, folding to 690 x 450 x 300 mm for transport. Body mass is approximately 35 kilograms with battery. The foldable design allows the robot to be stored in a standard large equipment case and transported by a single operator to laboratory spaces, demonstration venues, and conference sites without specialized vehicle or equipment.

The aluminum alloy structural components and ABS/PC blend exterior panels provide industrial-grade durability, while the full joint hollow design — routing all power and communications internally — keeps the external profile clean and reduces snag risk in collaborative research and demonstration environments.

43 Degrees of Freedom: Full G1 Ultimate Body

The G1 EDU Ultimate A's 43-DOF total includes the complete Ultimate tier body configuration:

Legs: Six degrees of freedom per leg (bilateral, 12 total), with extra-large joint ranges confirmed by Unitree: hip pitch ±154°, hip roll -30° to +170°, hip yaw ±158°, knee 0° to 165°. These ranges enable the G1's documented athletic capabilities including world-record 1.4-meter long jump, side flip, and kip-up from ground.

Arms: Seven degrees of freedom per arm (bilateral, 14 total) including wrist pitch ±92.5° and wrist yaw ±92.5° — the wrist DOF that enables the Dex3-1 hands to approach objects from any orientation within the arm's workspace.

Waist: Three degrees of freedom (yaw ±155°, pitch, and roll) enabling torso rotation, forward-backward lean, and lateral tilt for extended reach and whole-body balance coordination during manipulation tasks.

Dex3-1 Hands: Seven active degrees of freedom per hand (bilateral, 14 total) — thumb (3 DOF), index (2 DOF), middle (2 DOF). Added to the 29-DOF body of the EDU Plus tier, the 14 hand DOF bring the total to 43.

Dex3-1 Three-Finger Force-Controlled Hands

The Dex3-1 is Unitree's proprietary three-finger dexterous hand. Unitree's official product page confirms the kinematic distribution: The thumb has 3 active degrees of freedom; the index finger has 2 active degrees of freedom; the middle finger has 2 active degrees of freedom.

The three-finger layout covers the grasping repertoire most commonly required for research manipulation tasks:

Precision pinch grasp: Thumb opposed to index finger with the middle finger providing stability, enabling precision grasping of small objects between two fingertips.

Power grasp: All three fingers closing around an object for secure hold during transport, with maximum contact area for robust grip under dynamic conditions.

Cylindrical grasp: Three fingers wrapping around cylindrical objects (bottles, handles, tube-shaped components) with consistent contact across the finger pads.

Lateral key grip: Thumb positioned laterally alongside the index finger for key, card, and flat object handling.

Tripod pinch: All three fingers meeting at a central point for three-contact precision grasping of intermediate-sized objects.

Force-Position Hybrid Control: The Dex3-1 operates under force-position hybrid control — a control mode that blends position targeting with force limits, enabling the hand to close to a grasp configuration (position control) while maintaining force limits that prevent over-pressuring a grasped object (force control). This mode is what makes the Dex3-1 suitable for handling a variety of objects without damaging them through excessive grip force.

The U3 configuration does not include the tactile sensor arrays that the U4 adds — the Dex3-1 in the U3 uses joint-level motor current and torque estimation for contact force feedback rather than distributed surface-level tactile sensing. For research tasks where object fragility and slip detection do not require distributed tactile sensing, the U3's motor-torque-based force control provides sufficient contact awareness for reliable manipulation.

Technology and Specifications

Full G1 EDU Ultimate A (U3) Specifications

NVIDIA Jetson Orin NX 16GB: 100 TOPS for Manipulation AI

The Jetson Orin NX 16GB provides 100 TOPS of GPU-accelerated AI compute for running visual manipulation policies, the UnifoLM multimodal model, and custom AI systems developed through the SDK. For manipulation research, the relevant AI workloads include:

Visual grasping models: Detecting object position and orientation from camera data to plan grasp approach trajectories, running at the image processing rates needed for reactive manipulation.

Policy network inference: Executing imitation learning or reinforcement learning policies trained to map sensor observations to joint commands, requiring the GPU acceleration that the Jetson Orin provides.

UnifoLM multimodal interaction: Running the onboard UnifoLM model for voice command interpretation and image-based task instruction — enabling natural-language-commanded manipulation task execution.

The 16GB memory configuration supports the larger policy networks that manipulation research increasingly requires. BotInfo.ai's documented research workflow confirms the full development pipeline: Isaac Sim training with G1 URDF → domain randomization → Jetson Orin deployment via unitree_sdk2 → real-world iteration with VR teleoperation for imitation learning data.

UnifoLM-VLA-0 Open-Source Manipulation Baseline

The March 2026 open-source release of UnifoLM-VLA-0 (built on Qwen2.5-VL-7B, covering 12 manipulation task categories) provides U3 buyers with an immediately deployable manipulation AI baseline. This baseline provides working grasping and manipulation policies for common task categories — enabling research programs to start from a validated manipulation capability foundation rather than training from scratch.

The open-source release is particularly significant for the U3 because manipulation is the primary new capability that the U3 adds over the U2: having a working manipulation policy baseline available on day one allows research teams to evaluate baseline performance, identify limitations, and focus their research on extending or improving specific capabilities rather than on the complete training pipeline setup from scratch.

Applications and Use Cases

General-Purpose Manipulation Research

The U3's Dex3-1 hands enable research on object grasping, transport, placement, and basic assembly with a three-finger force-controlled end-effector. For manipulation tasks involving rigid objects in structured workspace configurations — standard laboratory objects, tool handles, boxes, containers, and components — the Dex3-1's force-position hybrid control provides the contact force management needed for reliable grasping without requiring distributed tactile sensing.

Published G1 research includes VR-teleoperated data collection for imitation learning, PPO-trained arm control policies deployed from simulation, and the SHIELD safety framework tested on the G1 in unstructured outdoor environments at Caltech, UT Austin, and Cornell. These research programs provide publicly available baselines that U3 buyers can build on rather than starting from zero.

Whole-Body Loco-Manipulation

Combining bipedal locomotion with manipulation is one of the most actively studied challenges in humanoid robotics. The U3's 43-DOF full-body configuration — with 6-DOF legs, 3-DOF waist, 7-DOF arms, and Dex3-1 hands — provides the physical substrate for research on whole-body loco-manipulation: tasks where the robot must walk, maintain balance, and simultaneously manipulate objects. The extra-large joint ranges of the EDU Ultimate body enable the whole-body motion coordination that manipulation during dynamic locomotion requires.

Pick-and-Place Logistics Applications

For logistics research involving standard package handling, the Dex3-1's power grasp capability combined with the G1's 3-kilogram arm payload covers the weight range of standard laboratory and light commercial packages. Research on autonomous pick-and-place — where the robot perceives an object, plans a grasp, grasps, transports, and places it — uses all of the U3's components: LIVOX LiDAR and RealSense cameras for perception, Jetson Orin for policy inference, and Dex3-1 for physical manipulation.

Human-Robot Interaction with Object Handover

Object handover — handing an object from the robot to a human or from a human to the robot — is a fundamental physical HRI task. The U3's Dex3-1 hands provide the grasping capability needed for handover tasks, and the force-position hybrid control regulates contact forces during the handover event. For HRI research programs that include handover as a study task, the U3 is the minimum G1 EDU configuration providing the necessary hardware.

Competition Robotics (RoboCup Humanoid League)

Unitree offers a G1-Comp configuration for competition use, specifically mentioning the RoboCup Humanoid League. The G1 EDU Ultimate A's 43-DOF configuration and full athletic body capability make it suitable for competitive humanoid robotics events requiring both locomotion agility and manipulation capability. Research teams entering robot competitions can use the U3's SDK for custom competition policy development.

Comparison: U3 vs. U2 vs. U4

The U3 is the correct choice for research programs that:

• Require physical object grasping and manipulation capability
• Do not specifically require distributed tactile sensing for their research objectives
• Are working within a budget that makes the $1,985 savings over the U4 meaningful
• Plan to use the UnifoLM-VLA-0 manipulation baseline as a starting point

Frequently Asked Questions (FAQ)

What is the Unitree G1 EDU Ultimate A (U3)? The Unitree G1 EDU Ultimate A (G1EDU-U3) is the entry-level dexterous manipulation configuration in the G1 EDU Ultimate tier, providing 43 total degrees of freedom with dual Dex3-1 force-controlled three-finger dexterous hands (7 DOF per hand: thumb 3 DOF, index 2 DOF, middle 2 DOF) without tactile sensors. It includes NVIDIA Jetson Orin NX 16GB (100 TOPS), an 8-core CPU, LIVOX MID-360 3D LiDAR, Intel RealSense D435i, dual encoders per joint, 120 N·m knee torque, 3 kg arm payload, 9,000 mAh quick-release battery (~2 hours), and full Python/C++/ROS 2 SDK. Priced at $64,292 and shipping in approximately two months.

What is the Dex3-1 hand, and what can the U3 manipulate? The Dex3-1 is Unitree's proprietary three-finger force-controlled hand with 7 active degrees of freedom (thumb: 3, index: 2, middle: 2) operating under force-position hybrid control. The U3 can grasp and manipulate rigid objects using power grasps (cylindrical objects, boxes, tools), precision pinch (small objects), tripod pinch, and lateral grip configurations. The arm payload of 3 kilograms covers standard laboratory and light commercial object weights. The Dex3-1 does not include tactile sensors in the U3 configuration — contact force feedback comes from joint torque estimation. For tasks requiring distributed tactile sensing, the U4 adds tactile arrays to the same hand.

What is the difference between the G1 EDU Ultimate A (U3) and Ultimate B (U4)? Both U3 and U4 feature identical 43-DOF body configuration with Dex3-1 three-finger hands. The only hardware difference is that the U4's Dex3-1 hands include integrated tactile sensor arrays, while the U3's Dex3-1 hands do not. This gives the U4 distributed contact pressure sensing for fragile object handling, slip detection, and haptics research — capabilities that the U3 cannot provide. The price difference is $1,985 ($64,292 for U3 versus $66,277 for U4). Research programs requiring only kinematic three-finger manipulation should consider the U3; programs requiring contact sensing should select the U4.

Where can I buy the Unitree G1 EDU Ultimate A (U3)? The G1 EDU Ultimate A (U3) is available through RobotShop (robotshop.com), OpenELAB Technology (openelab.io, 12-month warranty, 24-month EU/UK), RoboStore (robostore.com, includes RoboUniversity curriculum support), Top3DShop (top3dshop.com, buy or lease at $64,292), Futurology.tech, BotInfo.ai (North American institutional procurement, NET 30/60, grant documentation), and Roboworks (roboworks.net, EU/UK). Standard delivery is approximately two months.

Is the G1 EDU Ultimate A (U3) eligible for federal research grant funding? Yes. The G1 EDU configurations, including the U3, qualify as research equipment under DoD DURIP, NSF MRI, departmental equipment budgets, and other applicable federal research grant programs. BotInfo.ai provides grant-ready quotes and sole-source justification templates for institutional buyers navigating university procurement requirements. Grant narrative budget justifications should note the U3's specific research relevance: 43-DOF full-body kinematics for locomotion and manipulation research, Dex3-1 dexterous hands enabling three-finger grasping policy development, NVIDIA Jetson Orin 100 TOPS for on-device embodied AI model inference, and full ROS 2 SDK for integration with the academic robotics software ecosystem.

Summary

The Unitree G1 EDU Ultimate A (G1EDU-U3) represents the most accessible entry point into dexterous manipulation research on the G1 humanoid platform — providing 43 total degrees of freedom, dual Dex3-1 three-finger force-controlled hands enabling power, precision, and cylindrical grasping, NVIDIA Jetson Orin NX 16GB (100 TOPS) compute, dual encoders per joint, and complete Python/C++/ROS 2 SDK access at a confirmed price of $64,292. For research programs whose manipulation study objectives can be addressed with three-finger force-position hybrid control without requiring distributed tactile sensing, the U3 provides the complete manipulation research substrate — at $1,985 less than the tactile-sensor U4 and within an established platform ecosystem of 5,500-plus shipped units, 30-plus published research papers, and the UnifoLM-VLA-0 open-source manipulation baseline — making it the most cost-effective first step into humanoid robot manipulation research available from any manufacturer at commercial production scale.