Unitree G1 EDU vs Booster T2 Pro

G1 EDU does not have a single fixed bill of materials. Unitree lists 23-43 degrees of freedom (DoF), including hand options. The XR repository lists 23- and 29-DoF bodies; SDK examples cover 5- and 7-joint arms. Three-axis waist, seven-joint arms, hands and compute must be checked against the order. [S1] [S6] [S7]

T2 Pro has 31 body DoF: head 2, each arm 7, waist 3 and each leg 6. P1 has cosmetic zero-DoF hands, no wrist camera and no end-effector power/communication wiring. P2 uses grippers; P3 uses dexterous hands. Arm-end camera support is listed for P2/P3. Booster reports no standard P1-to-P3 retrofit path. [S3] [S10]

Unitree XR lists Dex1, Dex3-1, Inspire and BrainCo adapters. This establishes published software scope; the quotation must identify the supplied hand, wiring and sensors. [S7]

Booster identifies the standard P3 hand as BrainCo Revo 2 basic: 383 g per hand excluding the wrist, without tactile sensing. For P3-3 it reports Revo3 U21T, 21 active DoF, backdrivability, whole-hand tactile sensing, 0.01 N resolution and 70 N grip force. These figures come from correspondence because the hand document was inaccessible. Sensor resolution does not specify accuracy; grip force cannot be converted into robot payload. [S3] [S9]

T2’s 10 kg dual-arm rating is static and includes the hands. Subtracting the stated Revo 2 masses gives 10 - (2 × 0.383) = 9.234 kg. This is a mass calculation only. Wrists, adapters and arm posture are unaccounted for, so it cannot specify net object capacity. Booster recommends keeping loads near the torso; no posture/load curve was supplied. [S3] [S10]

G1 EDU lists approximately 3 kg arm load, dependent on posture. The definitions do not support a direct ratio against the dual-arm static figure or a 6 kg walking capacity by adding two G1 arms. G1 EDU knee peak torque and T2 maximum joint torque also refer to different joints and duty conditions; they do not establish a lifting ranking. [S1] [S10]

Booster’s final reply states up to 5 kg at up to 1 m/s in Walking mode, with upper-body control enabled or disabled. It states that walking with 7-8 kg is currently unachievable, citing waist control and motor heating. [S3]

The described tray demonstration lasts roughly 90 seconds and is operator-assisted. The 5.3 kg includes the tray; eleven water bottles account for approximately 4 kg according to the reply. The tray rests on the forearms rather than being gripped. We could not inspect the video. No hand-gripped dynamic load measurement or 20-30 minute thermal result was supplied. There is no G1 EDU measurement under the same task conditions in this review. [S3]

Booster’s corrected reply says arm commands through UpperBodyCustomControl(true) can run alongside factory walking/balance control. The inspected SDK lists T2 for this call. Full CUSTOM mode transfers every joint, including the legs, to customer code; the safety manual describes support fixtures for this mode. [S3] [S4] [S11]

ArmController’s joint map does not match T2. Source and the official helper guide use index 6 for right shoulder. T2’s 31-joint table maps 6 to left wrist and 9 to right shoulder. The class handles only indices 2-9 and creates a 23-element command array; it is unsuitable for T2 without modification. G1’s 5/7-joint rt/arm_sdk examples compiled in this review but were not run on physical EDU hardware. [S5] [S6] [S10] [S13]

Cartesian calls are marked K1/T1. However, the capability matrix also says some hand planners do not reject by model name and depend on factory software configuration. This does not prove failure on every T2. A runnable T2 example, firmware and RPC results are required; the source describes 501 for disabled capability and 502 for a motion unreachable from the current state. The general DDS guide’s CUSTOM warning for direct rt/joint_ctrl and its relationship with the upper-body path have not been checked on the device. [S4] [S14] [S16]

Booster identifies DDS rt/joint_ctrl for low-level torque commands. The source contains kRevoHand; Revo 2 support is stated in the reply. Hand angle/force scaling and firmware compatibility have not been verified. [S3] [S8]

The final reply states that there is no dedicated collision-protection API. PROTECT/DAMP functions concern joint limits and motor states. The sources do not establish validated collision prevention for people or objects in the workspace. [S3]

T2 Pro lists Jetson Thor T5000, 128 GB RAM and 512 GB storage. Its 2,070 TFLOPS figure specifies FP4 sparse conditions. G1 EDU offers different compute modules, including Orin; module, memory and storage depend on the order. TFLOPS/TOPS figures at different precisions cannot specify a task-speed ratio. [S1] [S10]

Booster says customers must deploy a suitable VLA model for grasping and releasing. The manual’s basic motion, chat and dance agents do not cover this task. No firm release date was given for dual-hand grasping examples, manipulation baselines, teleoperation or datasets. Unitree XR publishes teleoperation, recording and simulation workflows; Unitree also says some showcased functions remain in development/testing. No comparable model latency, memory-use or task-success measurements were available. [S1] [S2] [S3] [S7]

Booster reports JetPack 7.1-b112, Tegra R38 revision 4.0, Ubuntu 24.04.4 LTS / ARM64, ROS 2 Kilted, CUDA 13.0, cuDNN 9.12.0.46-1 and TensorRT 10.13.3.9-1. These versions come from the manufacturer reply; device output and Isaac ROS/Isaac Lab compatibility results were not inspected. [S3]

The inspected Booster SDK README uses Ubuntu 22.04 as a development-host reference. It is not a package-compatibility test for the robot image. For either robot, check the delivered image, hand driver, SDK revision and model through a build/run record on the device. [S4]

T2’s main battery is 48 V / 10 Ah. The magnetic battery, cable and adapter are reported as standard delivery items. Booster withdrew the earlier 1100 mAh, 5-12 V, 43 g and 4-5 hour figures as belonging to another product. Nominal energy is 480 Wh; no usable-energy measurement was supplied. [S3] [S10]

T2 states two hours of continuous walking; standing manipulation runtime is unmeasured. Operation while connected to charging is supported. Battery replacement requires shutdown, with roughly three minutes reported until ready again. G1 EDU lists a 9 Ah quick-release battery and approximately two hours; quick-release does not establish hot-swap support. Neither model has a duty-cycle test establishing 1.5 hours of manipulation in these sources. [S1] [S3] [S10] [S12]

Energy calculation: 480 Wh / 1.5 h = 320 W average power. This assumes all nominal energy is usable; cutoff, reserve and temperature may reduce the actual budget. It is not a runtime measurement. T2 SDK includes a rt/battery_state voltage/current/SOC example, which compiled. Actual measurement needs timestamped V/I/SOC, joint temperature, load and task phase recorded together. After device current sign and scaling are verified, consumption can be calculated in Wh. [S10] [S15]

G1 Basic pricing does not cover EDU; T2 P1 pricing does not cover P3. Match robot/hand codes, sensors, compute, software access, spare battery, delivery terms, taxes, training and commissioning. The correspondence quotes an extra T2 10 Ah battery at 900 USD FOB China; it is not a Turkey delivered price and needs a current quotation. [S3]

Unitree lists an 18-month EDU warranty and refers to its warranty brochure for terms. Check local contractual coverage, fall/misuse exclusions, spare parts, firmware access and service turnaround separately. An equivalent local T2 service contract was not supplied for this review. [S1]