PLEN
Contact us

Research platforms · Robot learning · Grants · Procurement

For universities and public buyers

Unitree EDU robots for research in locomotion, manipulation and robot learning: joint-level access, ROS 2, simulation models of the same robots, and the documents you need for grant applications and public procurement.

Discuss a research configuration Tender documents
  1. How we bid
  2. EDU platforms
  3. How robots are trained today
  4. Literature
  5. For courses: our training product
  6. Grant applications
  7. Procurement
Ink drawing: a clockwork robot scholar in a mortarboard points at a book on a lectern and looks at the moon through a telescope; a small tin mouse-robot carries a scroll.

How we bid and what we deliver for a tender

For universities and public buyers — four points a typical Unitree EDU platform offer covers. NeoHumanoidX is an independent integrator; we supply genuine Unitree robots purchased from Unitree’s authorised distributor in Poland. We are not an authorised Unitree dealer.

Delivery and commissioning

Delivery of the agreed configuration (model and EDU version, options), on-site commissioning, and team instruction. The scope is set out in the written offer.

Documentation in Polish

A Polish statement of requirements and tender document pack (including catalogue sheet, operating manual, warranty and service terms, company registry data). Details: Procurement.

Training platform

We are developing a training platform with an in-browser robot simulator, intended for courses with Unitree humanoid robots. See For courses.

Local service

Contact and support from Warsaw; warranty and service-response terms as stated in the offer for the given model. Service is agreed within our integration offer.

Amounts and lead times are given only in the written offer, not on this page. Request an offer

1. EDU platforms for research

For research you need the EDU variant. In Unitree's own specification tables, secondary development is available for G1 and H2 only in the EDU versions. In the Go2 family, AIR and PRO do not have it, Go2 X has basic support and Go2 EDU has full support. Secondary development is what gives your code access to the robot: joint-level commands and state, sensors and extra compute.

What EDU adds over the base version. Technical data as published by the manufacturer (unitree.com, accessed 22.10.2026)
G1 and G1 EDUGo2 AIR, PRO, X and Go2 EDUH2 and H2 EDU
Secondary development (SDK)base: no; EDU: yesAIR and PRO: no; X: basic support; EDU: fullbase: no; EDU: yes
Extra computebase: none; EDU: additional high-performance compute moduleAIR, PRO and X: none; EDU: additional compute module (40–100 TOPS)base: none; EDU: PC2 computer (Intel Core i7) for custom development and an optional high-performance compute module
DoF and handsbase: 23; EDU: 23 to 43, with an optional 2 waist DoF and an optional Dex3-1 hand (7 DoF) with 2 wrist DoF12 joint motors on all versions31 on both; EDU adds dexterous hands in several variants
Sensorsdepth camera, 3D LiDAR and a 4-microphone array on bothfoot force sensors and depth camera on EDU onlysame on both: stereo camera, microphone array and speaker
Othermaximum knee torque 120 N·m (base: 90 N·m); maximum arm load about 3 kg (base: about 2 kg)15,000 mAh battery (other versions: 8,000 mAh) and a fast charger—
Warranty18 months (base: 8)12 months (AIR: 6; PRO and X: 12)12 months (base: 8)

Sources: unitree.com/g1, unitree.com/go2, unitree.com/H2, unitree.com/Dex3-1, accessed 22.10.2026. Values are approximate and depend on the variant and configuration; "—" means not stated. Unitree names no specific compute module ("such as …"), so we specify it in writing in the offer.

Why EDU suits research

Joint-level access

unitree_sdk2 (C++, DDS; Python in unitree_sdk2_python) ships low-level examples per robot: example/g1/low_level/, example/go2/go2_low_level.cpp, example/h2/low_level/. The low-level state carries per-joint position, velocity and acceleration; low-level commands set position, velocity and torque targets with gains. The SDK targets Ubuntu 20.04.

ROS 2 without wrappers

unitree_ros2 uses the robot's DDS messages directly (CycloneDDS 0.10.x), so ROS 2 nodes talk to the robot without wrapping the SDK. Ubuntu 22.04 / Humble (recommended) and 20.04 / Foxy; low-level control examples for Go2/B2, H1-2 and G1.

User compute on board

Your code runs on compute separate from the platform computer that runs Unitree's own control: G1 EDU and Go2 EDU take an additional high-performance compute module (Go2: 40–100 TOPS), and H2 EDU adds a PC2 computer (Intel Core i7) for custom development and an optional high-performance compute module.

Interfaces

G1 / H2: Wi-Fi 6, Bluetooth 5.2. Go2 EDU: Wi-Fi 6 dual-band, Bluetooth 5.2; foot-end force sensors and depth camera only on EDU. H2 EDU back panel: USB-C / USB 3.0 to PC2 and two Ethernet segments (H2 User Manual V1.1).

Dexterous hands

Dex3-1: 3 fingers, 7 active DoF, 33 tactile (pressure) sensors, USB 2.0; feedback includes joint position, velocity, torque and temperature, pressure and IMU. Optional on G1 EDU. H2 EDU: several hand options. xr_teleoperate supports Dex1-1, Dex3-1, Inspire and BrainCo end effectors.

Simulation models of the same robots

URDF in unitree_ros (g1_description, go2_description, h2_description); MJCF in unitree_mujoco (G1 23/29 DoF, Go2, H2). unitree_mujoco exposes the same SDK2 low-level interface as the real robot, so a controller is checked in simulation before it runs on hardware.

Coverage in Unitree's training and teleoperation repositories

Repository / modelG1Go2H2
unitree_sdk2: low-level examples✓✓✓
unitree_ros (URDF) / unitree_mujoco (MJCF)✓✓✓
unitree_rl_gym: training in Isaac Gym, checking in MuJoCo, trial on the robot✓✓—
unitree_rl_lab: Isaac Lab✓ 29 DoF✓—
unitree_sim_isaaclab: manipulation and teleoperation tasks✓——
xr_teleoperate / unitree_lerobot✓——

Checked 22.10.2026; "—" = not listed. H2 is today covered by the SDK, URDF and MJCF only; learned whole-body control on H2 appears so far only in research preprints [P2]. Other models we supply: As2 (secondary development on X and EDU), A2 ("Supported", Intel Core i7 for user development) and B2 also have unitree_sdk2 examples; see Robots.

Licences. The software, models and datasets mentioned here are third-party projects (e.g. Unitree Robotics, Hugging Face), published by their authors and used under their own licences. Licences differ between projects, and between code, model weights and datasets. Please check the licence in each project before use. NeoHumanoidX does not grant any licence to them. xr_teleoperate and teleimager (Unitree Robotics): Apache License 2.0 according to the projects' LICENSE files; they build on third-party components under their own licences.

2. How robots are trained today

Current practice on Unitree-class robots combines a few building blocks: a locomotion or whole-body policy trained with RL in GPU simulation, manipulation learned from teleoperated demonstrations, and, increasingly, a pretrained vision-language-action model fine-tuned on the lab's own data. For each block: what it is, the main tools, and where to start reading. Where we work with the tools ourselves, the link goes to the relevant part of our R&D page.

  1. TrainRL in GPU simulation
  2. Randomisephysics, sensors, pushes
  3. Sim-to-simreplay in MuJoCo
  4. Sim-to-realstaged tests on the robot
  5. Fine-tunedemonstrations, VLA
2.1

RL in GPU simulation

Thousands of robot instances run in parallel on one GPU; an on-policy algorithm, usually PPO [R1], learns a policy from proprioception (optionally height maps or vision) to joint targets. legged_gym [P6] showed locomotion policies trained in minutes on a single GPU and is the foundation of unitree_rl_gym (with rsl_rl). Isaac Lab is the basis of unitree_rl_lab; MuJoCo [R15], including MJX on accelerators via JAX and MuJoCo Playground [R14], is the common alternative.

Tools: Isaac Lab / unitree_rl_lab · Isaac Gym / unitree_rl_gym · legged_gym · rsl_rl · MuJoCo / MJX
On our R&D page: Unitree research ecosystem

2.2

Domain randomisation

Physical and visual parameters are randomised during training so that the policy cannot overfit one simulator instance [P7] [R2]. Unitree's own G1 configuration in unitree_rl_gym randomises friction (0.1–1.25) and added base mass (−1 to +3 kg) and pushes the robot every 5 s. An alternative to pure robustness is online adaptation: RMA [R4] infers environment parameters from recent history.

Tools: domain_rand configs in unitree_rl_gym and Isaac Lab

2.3

Sim-to-sim, then sim-to-real

Before a policy touches hardware it is replayed in a second simulator: unitree_rl_gym documents a four-step path: train, play back, transfer to a second simulator (MuJoCo) and transfer to the robot (Train, Play, Sim2Sim, Sim2Real), and unitree_mujoco exposes the same low-level interface as the robot. Remaining gaps are closed with actuator models [R3] or learned corrections: ASAP [P1] trains a delta-action model on real G1 data.

Tools: unitree_mujoco · unitree_rl_gym deploy · unitree_rl_lab sim2real
On our R&D page: Unitree research ecosystem · the in-browser MuJoCo simulator we are developing for our training platform

2.4

Imitation learning from teleoperation

An operator drives the robot and the recorded episodes train a policy by behaviour cloning, e.g. ACT [R5] or Diffusion Policy [R6]. xr_teleoperate (Apple Vision Pro, PICO 4 Ultra Enterprise, Meta Quest 3; builds on Open-TeleVision [P8]) records G1 demonstrations, also in simulation with unitree_sim_isaaclab; unitree_lerobot converts them for training with LeRobot. Isomorphic, exoskeleton-style teleoperation on G1: [P4].

Tools: xr_teleoperate · unitree_sim_isaaclab · LeRobot / unitree_lerobot
On our R&D page: Unitree research ecosystem

2.5

Human video and cross-embodiment transfer

Teleoperation data is expensive, so recent work learns from human motion and video: retargeted human motion for humanoid teleoperation and imitation [R7] [R8], and egocentric human video with zero-shot transfer to G1 [P5]. Controllers are also transferred between robots (from G1 to H2 in [P2]), and cross-embodiment datasets pool data from many robots [R9].

Tools: research code accompanying the papers

2.6

Vision-language-action (VLA) and foundation models

VLA models map camera images and a language instruction to actions and are fine-tuned on a lab's demonstrations, for example π0 [R12], OpenVLA [R11] and RT-2 [R10]. Fine-tuning such models usually needs a powerful GPU.

Tools: LeRobot
On our R&D page: technology watch

We share fresh technical news from the Unitree ecosystem and robotics research on our LinkedIn page.

3. Literature: where to start

Core list (P) with the Unitree robot used, where the abstract names one; further reading (R) on the underlying methods. arXiv IDs checked against the stated titles on 22.10.2026.

These are third-party publications; their authors are not affiliated with NeoHumanoidX and do not endorse our products or services. Summaries are our own; see each paper for details.

Core papers

  1. [P1]T. He, J. Gao, W. Xiao, Y. Zhang, Z. Wang, J. Wang, Z. Luo, G. He, N. Sobanbab, C. Pan, Z. Yi, G. Qu, K. Kitani, J. Hodgins, L. Fan, Y. Zhu, C. Liu, G. Shi. "ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills." RSS 2025. arXiv:2502.01143 robot: Unitree G1
  2. [P2]Q. Lan, Z. Tian, Z. Wang, J. Liang. "TRACE: Efficient Transfer Learning for Neural Whole-Body Control Across Humanoid Embodiments." arXiv, 2026. arXiv:2610.17999 preprint robot: from G1 to H2
  3. [P3]Y. Lin, G. Shi, C. K. Liu. "Action-Aligned Latent Planning for Versatile Humanoid Loco-Manipulation (FIBER)." arXiv, 2026. arXiv:2610.24675 preprint robot: real G1: box pick-up, door opening, grasping
  4. [P4]Y. Wang, P. Yang, L. Zhuo, L. Wei, X. Du, K. Song, W. Wang, D. Ma. "Compositional Humanoid Manipulation through Isomorphic Teleoperation and Independent Skill Routing." arXiv, 2026. arXiv:2610.26169 preprint robot: Unitree G1
  5. [P5]Lambda-humanoid Team: C. Xu, J. Chen, J. Ye, et al.. "λ0: Scaling Open-World Egocentric Human Data for Humanoid Loco-Manipulation." technical report, arXiv, 2026. arXiv:2610.18950 preprint robot: zero-shot transfer to G1
  6. [P6]N. Rudin, D. Hoeller, P. Reist, M. Hutter. "Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning." CoRL 2021. arXiv:2109.11978 robot: basis of unitree_rl_gym (legged_gym)
  7. [P7]J. Tobin, R. Fong, A. Ray, J. Schneider, W. Zaremba, P. Abbeel. "Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World." IROS 2017, DOI 10.1109/IROS.2017.8202133. arXiv:1703.06907
  8. [P8]X. Cheng, J. Li, S. Yang, G. Yang, X. Wang. "Open-TeleVision: Teleoperation with Immersive Active Visual Feedback." arXiv, 2024. arXiv:2407.01512 robot: cited by xr_teleoperate

Further reading: methods

  1. [R1]J. Schulman, F. Wolski, P. Dhariwal, A. Radford, O. Klimov. "Proximal Policy Optimization Algorithms." arXiv, 2017. arXiv:1707.06347
  2. [R2]X. B. Peng, M. Andrychowicz, W. Zaremba, P. Abbeel. "Sim-to-Real Transfer of Robotic Control with Dynamics Randomization." arXiv, 2017. arXiv:1710.06537
  3. [R3]J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V. Tsounis, V. Koltun, M. Hutter. "Learning agile and dynamic motor skills for legged robots." Science Robotics 4(26), eaau5872, 2019. arXiv:1901.08652
  4. [R4]A. Kumar, Z. Fu, D. Pathak, J. Malik. "RMA: Rapid Motor Adaptation for Legged Robots." RSS 2021. arXiv:2107.04034
  5. [R5]T. Z. Zhao, V. Kumar, S. Levine, C. Finn. "Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware (ACT)." arXiv, 2023. arXiv:2304.13705
  6. [R6]C. Chi, Z. Xu, S. Feng, E. Cousineau, Y. Du, B. Burchfiel, R. Tedrake, S. Song. "Diffusion Policy: Visuomotor Policy Learning via Action Diffusion." arXiv, 2023 (extended version of the RSS 2023 paper). arXiv:2303.04137
  7. [R7]Z. Fu, Q. Zhao, Q. Wu, G. Wetzstein, C. Finn. "HumanPlus: Humanoid Shadowing and Imitation from Humans." arXiv, 2024. arXiv:2406.10454
  8. [R8]T. He, Z. Luo, X. He, W. Xiao, C. Zhang, W. Zhang, K. Kitani, C. Liu, G. Shi. "OmniH2O: Universal and Dexterous Human-to-Humanoid Whole-Body Teleoperation and Learning." arXiv, 2024. arXiv:2406.08858
  9. [R9]Open X-Embodiment Collaboration: A. O'Neill, A. Rehman, et al.. "Open X-Embodiment: Robotic Learning Datasets and RT-X Models." arXiv, 2023. arXiv:2310.08864
  10. [R10]A. Brohan, N. Brown, J. Carbajal, et al.. "RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control." arXiv, 2023. arXiv:2307.15818
  11. [R11]M. J. Kim, K. Pertsch, S. Karamcheti, et al.. "OpenVLA: An Open-Source Vision-Language-Action Model." arXiv, 2024. arXiv:2406.09246
  12. [R12]K. Black, N. Brown, D. Driess, et al.. "π0: A Vision-Language-Action Flow Model for General Robot Control." RSS 2025. arXiv:2410.24164
  13. [R14]K. Zakka, B. Tabanpour, Q. Liao, et al.. "MuJoCo Playground." arXiv, 2025. arXiv:2502.08844
  14. [R15]E. Todorov, T. Erez, Y. Tassa. "MuJoCo: A physics engine for model-based control." IROS 2012. DOI 10.1109/IROS.2012.6386109

We share fresh technical news from the Unitree ecosystem and robotics research on our LinkedIn page.

4. For courses: training platform (in development)

R&D

NeoHumanoidX training platform in Polish

We are developing a training platform with an in-browser robot simulator for teams that work with Unitree humanoid robots. The platform is not publicly available. Universities interested in a pilot can contact us.

  • modules with introductory lessons, ranging from safety around robots, through environment setup and communication with the robot (unitree_sdk2, DDS, ROS 2 Humble), motion control and reinforcement learning, Fast-LIO 2 odometry, to AI tools: LeRobot and Gemini Robotics-ER
  • each module: learning objectives, prerequisites, an introductory lesson, sources with licences, and an "as of" date
  • a 3D simulator in the browser based on MuJoCo (WebAssembly) with the public model of a full-size Unitree humanoid; no installation, no GPU server
Ask about a pilot Details of this work on the R&D page

5. Grant applications: the robot part

We help write the equipment part of applications that include a robot (NCN, NCBR, FENG, Horizon Europe): the rationale, the link to the research need, the parameters and the implementation costs. The science and the work plan are yours; we supply the platform-side facts, with sources.

Equipment rationale

Why this platform and this variant (usually EDU), what the research cannot do without it, which alternatives were considered and why they were rejected, and how the equipment will be used after the project.

From research need to parameters

A traceable table from each task or work package to the capability it needs, the specification that meets it and the manufacturer source. Example: whole-body reinforcement learning needs joint-level torque or position commands via the SDK, and dexterous manipulation needs the Dex3-1 hand (7 DoF, 33 tactile sensors).

Parameters with sources

Version, DoF, hands, sensors, user compute, interfaces, software stack and simulation assets, each with the official source and access date. Written so that it can later serve as the technical description in the tender.

Implementation costs (structure)

The cost items to budget: robot and options (hands, compute module, batteries), GPU workstation for training, teleoperation hardware, safety equipment and a test area, commissioning and staff training, warranty and service. We provide figures in a written quote, not on this page.

Risks and schedule

Delivery lead time, commissioning, the sim-to-real gap and safety, with mitigations such as staged tests (simulation first, then the robot without motion, then with motion).

Writing it so reviewers can score it

  1. Start from the evaluation criteria, not the template. Get the funder's guidance for reviewers and its scoring definitions for the call, and map every criterion to a specific paragraph. If it is not explicit and up front, a reviewer working through a large pile will not find it.
  2. Self-review against those criteria before submission. Go through the draft criterion by criterion and fix the gaps. An AI assistant can help an applicant with this check; first read the funder's rules on AI use and confidentiality.
  3. Use AI drafts and literature tools only as a starting point. Editing and disagreeing with a draft is faster than writing from zero, but every statement and every reference must be checked by the authors: generated text is often generic or overstated.
  4. Search for calls widely. Besides the usual national and EU databases, topic-matching tools can surface calls; always confirm eligibility and deadlines at the source.

6. Procurement

Tender document pack

Prepared for the specific model and version.

DocumentFormatAvailability
CE declaration of conformity (where the manufacturer issues one for the model)PDFon request from the contracting authority
Datasheet (PL)PDFon request from the contracting authority
User manual (PL)PDFon request from the contracting authority
Warranty card and service termsPDFon request from the contracting authority
Company registry extract (KRS)PDFon request from the contracting authority

A bid usually includes

  • price incl. transport, setup, instruction and licences
  • Polish product description with every parameter
  • delivery time for the chosen configuration
  • warranty and service with a response to reports within 72 hours

In the description of the subject of the order, state "EDU version with SDK access (secondary development)" explicitly.

Contact usAsk for price
This website uses only strictly necessary browser storage. More