Hardware Review

Unitree G1 Review 2026: Is It Worth It for Research?

The Unitree G1 arrived as one of the most price-competitive full-size humanoids on the market. After extensive testing at the SVRC Mountain View facility, here is what researchers and teams actually need to know before buying or leasing one.

Unitree G1 Specs: What You Get

The G1 stands approximately 1.3 meters tall and weighs around 35 kg in its standard configuration. It ships with 23 degrees of freedom distributed across the legs, torso, and arms. Each arm terminates in a three-fingered dexterous hand, giving it a meaningful manipulation payload of roughly 2 kg per hand. The onboard compute stack runs an NVIDIA Jetson Orin module, which is sufficient for running inference on mid-size vision-language-action models in real time. Battery capacity supports approximately 2 hours of mixed-use operation — locomotion plus light manipulation tasks — before requiring a charge cycle of about 1.5 hours.

Full Specification Table

Parameter	G1 Standard	G1 EDU (Research)
Height	127 cm	127 cm
Weight	35 kg	35 kg
Total DOF	23	23 (+ optional dexterous hand upgrade to 43)
Arm DOF (per arm)	7 (shoulder 3 + elbow 1 + wrist 3)	7
Leg DOF (per leg)	5 (hip 3 + knee 1 + ankle 1)	5
Waist DOF	1 (yaw)	1 (yaw)
Arm payload (per hand)	3 kg (close range), 1.5 kg (full extension)	3 kg (close range)
Walking speed (max)	2.0 m/s	2.0 m/s
Battery	9.9 Ah / 48V (475 Wh)	9.9 Ah / 48V (475 Wh)
Runtime (mixed use)	~2 hours	~2 hours
Charge time	~1.5 hours	~1.5 hours
Onboard compute	NVIDIA Jetson Orin NX 16GB	NVIDIA Jetson AGX Orin 64GB
Cameras	1x RealSense D435i (head)	1x D435i + 2x fisheye (head)
IMU	Built-in 6-axis	Built-in 9-axis
Price (2026)	$21,500 USD	$28,000 USD

You can view current pricing and availability through our Unitree G1 product page.

Joint Torque Specifications

The G1 uses Unitree's proprietary high-torque-density motors across different joint groups. Peak torques determine what manipulation and locomotion tasks are feasible:

Hip (roll/pitch/yaw): 88 Nm peak — sufficient for dynamic locomotion including running and jumping
Knee: 139 Nm peak — the highest-torque joint, enables stair climbing and recovery from stumbles
Ankle: 50 Nm peak — adequate for flat terrain walking, limiting factor for steep stair descent
Shoulder (pitch): 25 Nm peak — handles 3kg payload at close range, drops to 1.5kg at full arm extension
Elbow: 25 Nm peak — adequate for most tabletop manipulation tasks
Wrist (3-DOF): 8 Nm peak — sufficient for orientation control, limits heavy tool use

The torque budget for manipulation is the G1's most important practical constraint. At full arm extension (~0.5m from shoulder), the 25Nm shoulder motor supports approximately 1.5kg of payload after subtracting the arm's own weight. For data collection, this means keeping objects close to the body during manipulation — not a problem for tabletop tasks, but limiting for tasks that require reaching across a large workspace.

G1 vs H1: Which Unitree Humanoid?

Feature	G1	H1
Height	127 cm	180 cm
Weight	35 kg	47 kg
DOF	23	19
Payload (per arm)	3 kg	5 kg
Max speed	2.0 m/s	3.3 m/s
Dexterous hands	3-finger (upgrade available)	Not included (parallel jaw)
Price	$21,500	$90,000
Best for	Tabletop manipulation research, first humanoid	Human-scale tasks, industrial research

Choose the G1 if your primary use case is manipulation research on tabletop-scale tasks and you need the lowest entry cost. Choose the H1 if you need human-scale workspace reach, heavier payload, or if your research involves whole-body locomotion at high speeds.

ROS2 Integration and Python SDK

Unitree provides two development paths: a Python SDK for direct motor control and a ROS2 integration package for ecosystem compatibility.

# Python SDK: Direct motor control
import unitree_sdk2py
robot = unitree_sdk2py.G1()
robot.connect("192.168.123.15")  # Default G1 IP

# Read joint states
state = robot.get_joint_state()
print(f"Left arm joints: {state.left_arm_q}")  # 7-element array

# Command joint positions (position control mode)
robot.set_arm_joint_position(
    left_q=[0.0, -0.5, 0.0, 1.2, 0.0, 0.0, 0.0],
    right_q=[0.0, -0.5, 0.0, 1.2, 0.0, 0.0, 0.0]
)

# For data collection: record at 50Hz
import time
episodes = []
while collecting:
    state = robot.get_joint_state()
    images = robot.get_camera_frames()
    episodes.append({
        'qpos': state.q, 'qvel': state.qvel,
        'images': images, 'timestamp': time.time()
    })

The ROS2 package (unitree_ros2) publishes joint states on /joint_states, accepts commands on /joint_commands, and provides URDF and MoveIt2 configuration. Community-contributed packages on GitHub add teleoperation, motion capture retargeting, and LeRobot-compatible data recording.

Data Collection Assessment

The G1's suitability for building robot training datasets depends heavily on the task type:

Tabletop pick-place while standing: Good. The arm reach and payload handle most household objects. Throughput: 5-8 demos/hr with trained operator.
Mobile manipulation (walk + grasp): Adequate. Navigation is reliable in structured environments, but the base position uncertainty adds 2-3cm to grasp error. Budget extra demonstrations for robustness.
Bimanual coordination: Possible but limited. The 3-finger hands lack the dexterity of dedicated bimanual platforms like ALOHA. Suitable for coarse bimanual tasks (hold + pour) but not for fine bimanual assembly.
Locomotion data collection: Excellent. The sim-to-real pipeline for G1 locomotion is well-established with Isaac Lab environments. Collecting diverse locomotion demonstrations across terrain types is straightforward.

For manipulation-heavy data collection, consider pairing the G1 with SVRC's OpenArm 101 stations: use the G1 for mobile tasks and navigation demonstrations, and OpenArm for high-throughput tabletop manipulation data.

Locomotion Performance

Locomotion is where the G1 genuinely shines for its price class. In flat-floor environments, it achieves a comfortable walking speed of around 1.5 m/s and handles minor obstacles such as door thresholds and low floor lips without issue. Stair climbing works reliably on standard residential stairs when using Unitree's built-in locomotion controller; steeper industrial stairs require more careful tuning. Outdoor performance on concrete and packed gravel is solid, though loose terrain and wet surfaces expose the limits of the default foot contact planner.

Compared to the Go2 quadruped, the G1's locomotion robustness is — predictably — lower in unstructured outdoor environments. However for structured indoor settings such as warehouses, labs, and office buildings, the G1's bipedal gait is more than adequate. SVRC's internal benchmarks are available at the benchmarks page for direct comparison with other platforms.

Manipulation Capabilities and Limitations

This is where researchers need to calibrate their expectations carefully. The G1's three-fingered hand can grasp cylindrical and prismatic objects reliably, but dexterous manipulation — in-hand rotation, fine pinch grasps on small objects, or compliant insertion tasks — remains challenging. The hand's finger joint resolution is lower than dedicated research hands like the Allegro, and there is limited tactile sensing in the base configuration.

For data collection and imitation learning on pick-and-place tasks, stacking, and object handover, the G1 performs well. For contact-rich assembly or tasks requiring sub-millimeter precision, you will want to supplement with better wrist instrumentation or consider a dedicated arm platform. Our platform comparison tool can help you weigh the G1 against purpose-built manipulation hardware.

Software Ecosystem

Unitree provides a ROS 2 SDK, a Python API, and a WebSocket-based teleoperation interface out of the box. The community around Unitree hardware has grown significantly through 2025 and 2026, with open-source locomotion controllers, motion capture retargeting scripts, and LeRobot-compatible data collection pipelines available on GitHub. Unitree's own developer portal hosts firmware updates, URDF models, and simulation environments for Isaac Sim and MuJoCo.

One practical note: the default SDK has some rough edges around arm-locomotion coordination. If you want the robot to manipulate while walking — rather than stopping to grasp — plan for additional integration work. SVRC's engineering team has built bridge modules for common research workflows; ask about these when you contact us.

Who the G1 Is For

The G1 is an excellent first humanoid for academic labs, pilot programs, and teams building imitation learning datasets who need a mobile manipulation platform but cannot justify a six-figure hardware budget. It is also well-suited for companies building humanoid software stacks who need a real physical platform for testing — the low acquisition cost means you can instrument it aggressively, crash it occasionally, and learn fast without catastrophic financial risk.

It is not the right choice if your primary need is high-precision manipulation, fully dexterous hand tasks, or extreme locomotion robustness. In those cases a dedicated research arm or a higher-specification humanoid from our hardware catalog will serve better.

Verdict

At its price point, the Unitree G1 is hard to beat for research teams in 2026. Locomotion is reliable in structured environments, the software ecosystem is maturing rapidly, and the onboard compute is genuinely usable for on-robot inference. Manipulation capability is the honest weak point, but it is sufficient for most imitation learning workflows. SVRC offers the G1 for direct purchase and for short-term lease if you want to evaluate it before committing. See the G1 product page to get started, or use our comparison tool to stack it against alternatives.