H1-2

Manipulation-focused research where H1 was underpowered

Original H1 shipped with 4-DOF arms rated for ~5 kg lift and stalled on any pick-and-place heavier than a stapler. H1-2 raised rated arm payload to 21 kg per side and added a full 7-DOF arm — matching adult human range of motion. Labs at CMU, MIT, Stanford, ETH Zurich, and Tsinghua re-purchased into H1-2 for contact-rich manipulation and warehouse-style bin transfer. Published work covers bin-pick, box-stack, and door-through walking — categories H1 could not touch. The Robot Report tracked 40+ university H1-2 orders by end of Q1 2025.

Foundation-model data collection at adult body proportions

Skild AI, 1X, and stealth groups bought H1-2 for whole-body motion data targeting adult-scale production hardware. G1 at 127 cm forced post-processing to correct proportion mismatch; H1-2 at 178 cm matches Figure, Apptronik Apollo, and Tesla Optimus envelopes. Manufacturer has not disclosed buyer names, but distributors confirmed multiple 10+ unit orders through 2025. Buyers frequently add the dexterous hand option for hand-scale motion capture.

Dexterous demos with 3-finger hands option

The $128,900 configured H1-2 adds Unitree's 3-finger dexterous hands (7 DOF each, ~20 DOF added system-wide). Published demos cover tabletop assembly, single-arm door opening, cup pouring, and light bimanual manipulation. Hand-scale sim-to-real transfer stays uneven, matching G1 EDU behavior — locomotion policies transfer reliably, contact-rich hand policies need on-hardware fine-tuning.

Loco-manipulation and walk-and-carry

H1-2 became the practical Unitree platform for whole-body loco-manipulation — arms strong enough to carry a real payload while walking. Published 2025 work includes walking-with-tray demos, box-carry between rooms, and door-through with tool in hand. Frontier problem for humanoid RL: locomotion RL and manipulation RL individually work, combining them under load does not. H1-2 is the shipping platform where labs actually test combined policies.

Commercial demo fleets at distributor level

Unitree distributors upgraded rental fleets from H1 to H1-2 through late 2024 for bookings expecting manipulation, not just walking. Day rates $5,000-$9,000 depending on region and operator. Bookings cover trade shows, marketing shoots, and product launches. H1-2 has not shipped into commercial production duty at any warehouse or factory customer, per Unitree.

H1-2 initial launch — August 2024

27 DOF, 178 cm, 70 kg, starting price $99,900. Ships with 7-DOF arms per side (up from H1's 4), 21 kg rated arm payload, dual Intel i5 + i7 compute with optional Jetson Orin NX. Sensor stack (Livox MID-360, RealSense) matches H1 EDU. Unitree explicitly positioned H1-2 alongside the original H1 rather than as a replacement — H1 stays cheaper for locomotion research, H1-2 targets manipulation labs.

Dexterous hands configuration — October 2024

3-finger dexterous hands option pushed configured price to $128,900. Each hand adds 7 DOF, bringing total system DOF over 40 in the fully-optioned build. Targets manipulation-oriented research and foundation-model data collection. Hand controllers ship as a Unitree first-party module rather than a third-party retrofit.

ROS2 parity with H1 — Q4 2024

First-party ROS2 Humble packages inherited from H1 and G1 rollouts. Coverage: joint-state, IMU, LiDAR, depth topics, plus URDF and Gazebo/Isaac Sim descriptions. H1-2-specific URDFs shipped through Unitree GitHub. Third-party MuJoCo Playground added H1-2 model in Q1 2025.

Firmware 2.1 — loco-manipulation baseline — Q1 2025

Firmware update added a manipulation-aware balance controller — sustained walking with a 5 kg arm load without gait destabilization. Rough-terrain gait remained at 2.0 m/s sustained. Unitree did not publish a sprint controller — the extra mass from arms and torso rules out H1's 3.3 m/s record on H1-2 hardware.

Loco-manipulation benchmark inclusion — Ongoing through 2026

By early 2026 H1-2 appears in CMU StaRLING-Manip, Stanford Sim2Real-Manip, and Unitree's own Playground manipulation benchmarks. Locomotion RL transfers reliably; loco-manipulation transfer improved from Q4 2024 baselines but remains the hardest sim-to-real category on the platform.

Unitree HQ — Hangzhou, China

Primary R&D and demo facility. H1-2 launch footage, manipulation demos, and firmware 2.1 validation runs originate here. Internal fleet size not disclosed.

Source: https://www.unitree.com/h1/

University manipulation labs — global

40+ H1-2 units shipped into university labs through Q1 2025, per The Robot Report. Named institutions include CMU, MIT, Stanford, ETH Zurich, Tsinghua Advanced Robotics. Most orders come from labs upgrading from H1 or adding H1-2 alongside G1.

Source: https://www.therobotreport.com/

Foundation-model data-collection fleets

Skild AI, 1X, and stealth-stage groups placed multiple 10+ unit H1-2 orders through 2025 per distributor confirmations. Buyer identities and fleet sizes not released. Adult-scale body proportions make H1-2 the preferred Unitree platform for foundation-model motion data.

Source: https://www.unitree.com/

Distributor demo and rental fleets

Unitree USA (San Diego), Europe (Frankfurt), and Japan operate rotating H1-2 demo units. Day rates $5,000-$9,000 including operator. Replaced H1 as the primary demo unit for manipulation-oriented bookings through late 2024.

Source: https://www.unitreerobotics.eu/

70 kg mass creates indoor safety envelope

H1-2 weighs 70 kg — 23 kg more than H1 and heavy enough that an unplanned fall around bystanders is a real safety incident. Universities enforce operator-only clearance zones and a fall-arrest tether during dynamic trials. Contrast G1's 35 kg envelope where crash-and-recover is an equipment problem, not a safety incident.

Speed drops from H1's record to 2.0 m/s

H1 held a 3.3 m/s sustained-run record. H1-2's added arm and torso mass rule out the same envelope — Unitree has not published a sprint controller and has explicitly said the mechanical envelope does not support one at 70 kg. Sustained walking sits at 2.0 m/s. Locomotion-only research groups stayed on H1 or moved to G1 rather than take the H1-2 speed hit.

Runtime drops under manipulation load

864 Wh battery matches H1 — but H1-2 pulls significantly more power under arm load. Nominal 120-minute runtime shortens to 60-80 minutes during sustained manipulation trials per lab feedback. No hot-swap. Continuous shift work stays outside the design envelope.

No IP rating, no outdoor duty

H1-2 is not IP-rated. Rain, snow, dust, sub-freezing operation outside spec. All published outdoor footage runs on dry pavement in mild weather. Manufacturer has not disclosed a weatherproofed SKU.

Successor risk on multi-year amortization

H2 rumors circulated through late 2025 and internal BiorobotAI tracker residuals put H1-2 at 65% after year one, falling to 30% by year three. Buyers financing over 3+ years should price faster depreciation than the $99,900 headline suggests.

vs. Unitree H1

Same Unitree SDK2, same ROS2 packages, same sensor stack in top configurations. H1 keeps the 3.3 m/s sustained-run record; H1-2 loses it to the extra mass. H1-2 gains 7-DOF arms, 21 kg payload, and adult-scale manipulation envelope. Labs picking between them ask whether their study is locomotion-primary (H1 is cheaper and faster) or manipulation-primary (H1-2 is the only option). Unitree sells both in parallel through 2026.

vs. Fourier GR-1

Chinese adult-scale manipulation humanoid at 165 cm, 55 kg, 40 DOF. Fourier ships its own SDK and does not have Unitree's ROS2 first-party ecosystem or the university-lab footprint that Unitree built through 2024-2025. GR-1 has stronger integrated hand dexterity out of the box. H1-2 wins on ecosystem maturity, tooling, and community — GR-1 wins on DOF count and hand hardware at similar price.

vs. Agility Digit

Digit ships into commercial warehouses (GXO, Amazon pilots) as Robots-as-a-Service at ~$30/hr. Different segment from H1-2's research and data-collection sales. Digit's digitigrade legs, self-charging battery, and ISO 13482 certification target 24/7 industrial duty; H1-2 targets research where openness and sensor stack matter more than certification. Buyers picking between them are asking whether the problem is a research study or a shipping warehouse workflow.