Chapter 4: Humanoids and Quadrupeds — The General-Purpose Body Race

Overview

This chapter asks a deceptively simple question. Among China's humanoid and quadruped companies, which firms are building repeatable products, developer ecosystems, production capability, and field-learning loops—not merely the most impressive motion demo? A defensible answer is not a one-line ranking. Locomotion, manipulation, software openness, production, shipment, customer acceptance, and operating deployment must remain separate axes.

The thesis is that the general-purpose-body race will not be won by the company with the most spectacular demonstration. It will be won by a company that can operate a closed loop from data collection through policy learning, safety stops, maintenance, and renewed learning while respecting the hardware's limits. Humanoids may reuse spaces and tools designed for people, but they must solve bipedal balance and two-handed contact together. Quadrupeds have a relative advantage in terrain traversal and stability, but useful work often requires an arm, gripper, and additional sensing. The two forms are platforms with different failure costs, not simple substitutes heading toward the same finish line.

AgiBot, Unitree Robotics, and UBTECH Robotics are the locked deep profiles in this chapter. This selection does not declare them an absolute top three. It provides three contrasting evidence structures: portfolio breadth, developer access, and production or issuer disclosure ^{[1] [4] [7]}. Missing values are not estimated. Eligible media-corpus counts remain unmeasured. The official WRC 2026 exhibitor directory verifies Unitree at C215 and UBTECH at C103; AgiBot's booth and all three program entries remain unverified.

After reading this chapter... - You can explain why humanoids and quadrupeds carry different locomotion, manipulation, and safety burdens. - You can distinguish demonstrations, production, shipments, customer acceptance, and sustained deployment as different evidence stages. - You can compare AgiBot, Unitree, and UBTECH with the same product, technology, business, and ecosystem questions. - You can assess what open SDKs, ROS packages, and simulator assets do—and do not—contribute to factory integration. - You can define the task, logging, KPIs, safety ownership, and approval path for a general-purpose body in a manufacturing cell.

4.1 Five Evidence Rungs for Reading a General-Purpose Body

A robot video establishes that an action happened at least once. A buyer asks a different question: will it happen after the shift changes, lighting moves, the battery drains, and the joints heat up? The first evidence rung, a demonstration, confirms an action's existence but may omit the denominator: total attempts, failures, human interventions, resets, and edited takes. A live stream or continuous take reduces some editing ambiguity; it still does not establish weeks of uptime.

The second rung, a controlled evaluation, discloses a task, conditions, repetitions, and baselines. Peer-reviewed papers and public benchmarks belong here. HumanoidBench offers a common simulated language for whole-body locomotion and manipulation, yet a simulated benchmark cannot substitute for cables, vibration, contamination, or component variance on a factory floor ^[8]. A paper can bound an algorithmic claim well. It is weak evidence for company shipments, customer economics, or product support.

The third rung is production: finished units leave a manufacturing line. The fourth is shipment, sale, or revenue recognition: a product enters a transaction flow. The fifth is operating deployment: the system repeatedly produces value in a customer's task. These rungs may accumulate, but they do not follow automatically. A produced robot may be inventory, a research unit, an exhibition unit, or an internal data-collection device. A shipped robot may still await acceptance. Revenue may combine hardware, services, and operations. Before comparing “units,” one must read the unit definition and reporting period.

This ladder is not designed to dismiss corporate disclosure. It makes different achievements legible on fair terms. A young firm announcing a manufacturing milestone offers meaningful evidence. A listed issuer reporting revenue and volume in an audited report offers a different and generally stronger business record. The comparison breaks when a production count is plotted as if it were audited sales, or when a race record is added to factory uptime.

4.2 Choosing the Body: Two Legs, Four Legs, and Task Completion

The central value of a humanoid is not human-like appearance. It is the possibility of using environments built around human height, aisles, stairs, shelves, handles, and tools without redesigning all of them. That advantage is also a control burden. The support polygon is small, contacts change, and rapid arm motion reacts through the torso and feet. Once the robot grasps an object, walking and manipulation cease to be independent problems. They become one whole-body contact problem.

A quadruped can use a wider support base and more foothold choices to handle slopes, gravel, stairs, and disturbances. Where reaching a location and observing it constitute most of the task—power inspection or hazardous reconnaissance, for example—the form can be rational. Turning a valve, picking a component, or connecting a cable requires an arm and end effector. The added mass raises the center of gravity, consumes battery, expands the collision volume, and creates self-collision problems. The stability advantage must be tested again after adding manipulation hardware.

Classical whole-body control represents task priorities and joint redundancy in a model, coordinating several objectives across the body ^[15]. Recent learning systems absorb human motion, simulated experience, and real-robot data into policies. The approaches are not mutually exclusive. In a manufacturing cell, model-based constraints can enforce a safety envelope and joint limits while a learned policy handles observations and contact variability. A useful procurement question is not whether a product “has AI,” but which layers are deterministic, which are learned, and which can be independently stopped.

4.3 Deep Profile I — AgiBot

The English brand is AGIBOT or AgiBot, the Korean rendering used here is 애지봇, and the official Chinese operating name is 智元创新（上海）科技股份有限公司. Official material records a February 2023 foundation and a Shanghai manufacturing base ^[1]. The available evidence does not fully resolve every group operating entity or the legal scope of the term “headquarters.” The profile therefore records a Shanghai base while marking complete entity-and-headquarters resolution as partial.

Flagship products include the A-series full-body humanoids, the X series for open research and agile behavior, the task-oriented Genie line, and the D1 quadruped family. This breadth suggests a platform company spanning bodies, data, and development tools rather than a seller of a single humanoid. The official product and research pages connect AGIBOT World data, Genie Studio, simulation, teleoperation, and hands into a learning loop ^[2]. Portfolio breadth alone does not establish production volume, field uptime, or common software compatibility for every family.

The most distinctive technical choice is vertical integration of “embodiment plus AI.” Genie Sim 3.0 separates physics and rendering pipelines and targets large-scale synthetic data and automatic evaluation ^[3]. AgiBot's research material reports more than one million trajectories across more than 100 robots, five embodiments, and more than 1,000 scenarios ^[2]. These are issuer-disclosed dataset counts. Trajectory duration, duplication, quality filters, and success-to-failure balance are not audited under a cross-company standard. The numbers signal strategic direction but cannot support a dataset superiority ranking.

AgiBot reports that its 1,000th general-purpose embodied robot rolled off the production line in January 2025 ^[1]. This is a material issuer-disclosed manufacturing milestone, but it is not audited shipment volume, paid-customer count, or active deployment. “Rolled off,” “delivered,” “accepted,” and “operating normally” must stay separate for the milestone to be neither exaggerated nor dismissed.

The developer surface includes X1 design files, a bill of materials, assembly procedures, and training and inference code, along with open components such as AimRT. The captured evidence does not establish official ROS or ROS 2 support across the full portfolio, a validated NVIDIA Isaac Sim or Isaac Lab matrix, or a long-term support policy. “Open assets exist” is supported; “the entire portfolio is an open ecosystem” is not. Manufacturing integration still requires firmware versions, real-time-control privileges, safety interlocks, and commercial-use terms.

The planned official product photography should compare the whole-body joint layout of an A2 or X2 with the D1 quadruped structure. A photo documents visible hardware, not autonomy. Its company provenance and capture context should be explicit, and the image should not visually imply production or deployment that the caption cannot support.

4.4 Deep Profile II — Unitree Robotics

The Korean rendering is 유니트리 로보틱스, the English brand is Unitree Robotics, and the Chinese brand is 宇树科技. Official chronology describes the origin of the company in 2016, while the evidence assembled for this chapter does not resolve the current legal entity and headquarters address at registry level ^[4]. The foundation is recorded as 2016; headquarters remains unverified rather than being filled from common knowledge.

Unitree's distinctive position is a broad portfolio across humanoids and quadrupeds. H1, G1, and R1 are the principal humanoid bodies; Go2 and B2 cover consumer or research and industrial quadrupeds; the Z1 arm and Dex hands extend the manipulation surface ^[5]. A research team can study locomotion, teleoperation, and hand integration across bodies from one supplier. Breadth also multiplies firmware, communications, simulation, and safety-version combinations. A buyer should request a model-by-computer-by-operating-system support matrix, not accept a generic “supported” label.

Official material emphasizes internally developed motors, reducers, controllers, lidar, perception, and motion control ^[4]. Developer documentation lists SDKs for H1, G1, R1, Go2, B2, Z1, and Dex hands. Official open-source resources expose SDK2, Python interfaces, ROS and ROS 2 packages, Gazebo models, and reinforcement-learning examples ^[6]. This is a meaningful entry surface for low-level control, state access, and simulation. Official Unitree repositories now provide limited Isaac and Isaac Lab examples for Go2, H1, and G1, but they do not establish a portfolio-wide validated version matrix or long-term support contract. Example availability should not be mistaken for vendor-supported integration across the portfolio.

Unitree's official chronology and developer documentation establish a broad product and developer scope spanning humanoids, quadrupeds, an arm, and hands, but they do not provide audited shipment volume or field uptime ^{[4] [5]}. Stage performances, running, combat, and competition records are informative about dynamic control and hardware capability. Their denominators differ from intervention-free manufacturing hours and quality yield.

The paper ecosystem offers a separate form of evidence. A fast off-policy reinforcement-learning study trained locomotion on one RTX 4090 in roughly fifteen minutes and transferred policies to the G1 and Booster T1 ^[10]. Open-TeleVision combined H1 and hands in an immersive bimanual teleoperation and data-collection pipeline ^[13]. These papers show that Unitree bodies are accessible to external researchers. They do not mean the company warrants each research stack as a product feature, nor do they establish reliability outside the reported tasks.

The planned official photographs should show the G1 whole-body structure and a B2 or Go2 quadruped separately. A stage-performance image highlights agility but can conceal development ports, sensors, payload interfaces, and service access. For manufacturing readers, an official side view with visible joints and sensing is more useful than a hero image.

4.5 Deep Profile III — UBTECH Robotics

The Korean rendering is 유비테크 로보틱스, the English listed-issuer name is UBTECH ROBOTICS CORP LTD, and the Chinese legal name is 深圳市優必選科技股份有限公司. The 2025 annual report identifies a March 2012 establishment and headquarters in Shenzhen, China ^[7]. Among the three profiles, this filing connects legal identity, headquarters, finance, and product sales most strongly because audited statements and issuer disclosure are available.

The core humanoid line is Walker S, S1, and S2, with Walker S2 aimed at industrial handling, sorting, and inspection. The portfolio also includes the wheeled Cruzr S2, successive dexterous hands, and service and education robots. Technical disclosure combines high-performance servos, learning-based motion control, vision-language-action models, BrainNet 2.0, Co-Agent, Thinker models, and multi-robot coordination ^[7]. This points toward an industrial system strategy spanning hardware, data, and operational software rather than body sales alone.

The report describes Walker S2's autonomous battery exchange, whole-body degrees of freedom, reach and payload, and successive hand designs. It also says small-scale mass production and delivery began in 2025. The important distinction is among annualized capacity, production, sales volume, revenue, and active field units. Annualized capacity above 6,000 units is a rate-based capacity statement, not actual output or delivery. By contrast, disclosed 2025 sales volume of 1,079 units and approximately RMB820.6 million of revenue cover full-size embodied humanoid products and services for the reporting period ^[7]. The combined products-and-services scope must remain attached.

UBTECH reported total 2025 revenue of about RMB2.001 billion, gross profit of roughly RMB753.8 million, and a full-year loss of about RMB789.8 million ^[7]. Revenue growth and a smaller loss indicate commercialization progress, not completed profitability. Customer-level acceptance criteria, warranty burdens, service revenue per robot, and long-run uptime cannot be read from aggregate statements. Audited corporate numbers are a stronger comparison baseline; they are not a proxy for task reliability.

On ecosystem integration, the annual report describes in-house training, simulation, data-management, and fleet platforms, open Thinker model components, and internal use of NVIDIA Isaac for navigation simulation ^[7]. The captured evidence does not establish a general external SDK, official ROS or ROS 2 package matrix, or customer-facing Isaac deployment extension. UBTECH appears deep in internal integration. How much of each layer an external developer can replace and validate remains a due-diligence question.

The planned official photographs should prioritize Walker S2 handling containers in an industrial cell and the autonomous battery-exchange arrangement. A static stage pose offers less information than a scene containing workpieces, safety separation, and adjacent equipment. Even a good official factory image does not establish repeat success or intervention-free time and must remain separate from disclosed volume.

4.6 Comparing the Three Companies on the Same Scale

All three firms describe a union of whole-body robots and AI, but their evidence structures differ. AgiBot rapidly connects bodies, data, simulation, and open X1 assets. Unitree stands out for humanoid-and-quadruped breadth, public SDKs, and adoption by external researchers. UBTECH presents industrial tasks and comparatively concrete sales, revenue, and capacity disclosure through a listed issuer. The difference should change procurement weights; it should not be compressed into one composite score.

“Relatively attractive” does not mean technically superior. An external lab testing locomotion rapidly may value documentation and community adoption most. An automotive plant justifying capital to an investment committee may value audited disclosure, contractual acceptance, and support records. A team building its own data factory may value version-coherent teleoperation, simulation, hands, and policy deployment. The same robot has a different most-dangerous blank field under each objective.

4.7 From Simulation to the Floor: Two Different Reality Gaps

Humanoid learning results remain bounded by simulator, sensing, hardware, task, initialization, and reset protocol; they do not generalize automatically beyond those conditions ^{[8] [9] [10]}. A gait that succeeds in simulation meets motor delay, friction, battery-voltage sag, camera occlusion, and floor compliance in reality. Success on one physical unit does not ensure the same policy will run on the next unit with manufacturing variation.

System identification and motion retargeting address different sim-to-real failure modes ^{[11] [12]}. Identification reduces the difference between virtual and physical robot dynamics—motors, friction, bias, and delay. Retargeting transforms human motion and contact relationships into physically feasible trajectories for a robot with different link lengths and joint ranges. Fixing one does not remove error in the other.

PACE demonstrates a systematic route for fitting joint-level physical parameters and a global delay from short encoder traces ^[11]. OmniRetarget focuses on preserving interaction among the human, terrain, and manipulated object while constructing whole-body loco-manipulation trajectories ^[12]. Perceptive Humanoid Parkour joins depth perception, motion matching, teacher policies, and student distillation to chain dynamic skills on a G1 ^[9]. Together, they show both that a body is a vessel for learning and that camera field of view and hand capability bound the policy.

The manufacturing lesson is not merely to “use simulation.” It is to define which errors are estimated from which logs and when recalibration occurs. Does torque response change with motor temperature? Does slip increase as a finger pad wears? Does latency shift after firmware updates? A policy version without hardware state cannot reproduce a failure. Conversely, randomizing every variable can make a policy overly conservative or destabilize learning.

4.8 An Open Ecosystem Is Replaceability, Not Repository Count

Public repositories reduce startup friction. In manufacturing, openness has at least three levels. First, are observation and command data types, coordinate frames, cycles, and time synchronization documented? Second, do simulation and hardware use commands and constraints with the same semantics? Third, when firmware, SDKs, and models change, are compatibility, security fixes, and long-term support defined? Visible code and shared operating responsibility are not the same thing.

ROS or ROS 2 packages help message transport and tool integration, but they do not create hard real-time behavior, safety certification, or stable control. Gazebo, MuJoCo, and Isaac assets must go beyond matching joint names and appearance. Mass, inertia, transmission, contact, sensor noise, and control periods must reflect hardware closely enough for transfer. A buyer should request validated version combinations and regression results rather than ask only whether a tool is “supported.”

AgiBot's X1 release gives teams a deep entry point into hardware construction. Unitree's SDK and ROS resources aid rapid work across several bodies. UBTECH's filing emphasizes integration across internal data, training, simulation, and fleet operations. Each approach has value. None should be flattened into “fully open.” External replaceability, internal vertical integration, and commercial support are separate dimensions.

4.9 Safety: A System That Fails Safely

Humanoid safety cannot be reduced to gait success. When a human-height machine falls, risk includes body mass, a swinging arm, a carried object, and secondary impact with nearby equipment. A quadruped may look less threatening because its center of mass is lower, yet fast travel, an attached arm, and a slip at a stair edge create different hazards. The safety goal is not a claim that failure will never happen. It is detection, energy limitation, safe stopping, and recoverability by people.

A manufacturing cell should place an independent safety layer above the policy. Speed and force limits, geofencing, protective stop, emergency stop, communication-loss behavior, battery thermal faults, and fall direction require distinct tests. A learned policy interpreting safety sensor semantics is not sufficient as the only stopping path. Safety functions should operate separately from model updates and support change control and periodic proof tests.

When robots and people use the same aisle, “may be nearby” must remain distinct from “validated for collaborative work.” Even slow motion with a long arm or carried load can create a trapping hazard. Teleoperation does not solve safety automatically. Network latency, operator occlusion, authority transfer, and link loss become new hazards. Early deployment should begin with physical separation and observed operation, then admit only validated risks into shared space.

4.10 Manufacturing Walkthrough — Mixed-Tote Supply

Consider a task that finds totes at different shelf heights in a parts store and delivers them to a buffer beside an assembly cell. People occasionally enter the aisle. Tote handles vary in pose, and a shallow ramp and threshold interrupt the route. A humanoid may use existing shelf and handle heights. A quadruped may traverse the route and thresholds robustly. To finish the task, however, the quadruped needs a mounted arm or a handoff to a person or fixed robot.

Step one decomposes “move from store to cell” into task states: idle, receive mission, localize, travel empty, identify tote, approach, grasp handle, lift, carry, place, confirm handoff, charge or exchange battery, and recover. Every state needs an entry condition, success condition, timeout, and permitted human intervention. Demos often hide the restart location after failure, so acceptance tests must measure both a complete reset and local recovery.

Step two selects the body. If tote mass and handle pose are fixed and the floor is flat, a wheeled mobile manipulator may be simpler. If stairs and narrow human-oriented fixtures dominate, a humanoid can be justified. If observation and patrol dominate while manipulation is rare, a quadruped paired with a fixed arm may have better economics. The sequence should be reversed from “buy a general body and find a task”: begin with the hardest contact and terrain that actually demand a morphology.

Step three designs data and control. Teleoperation should collect handle occlusion, skewed totes, slip, human entry, degraded communications, and low battery—not only successful examples. Camera, depth, joint, torque, foot contact, hand contact, commands, and safety events need a shared clock. A high-level policy chooses the next state and target; a low-level controller enforces balance and force limits. A human operator recovers out-of-distribution states and labels those segments for new training.

Step four is acceptance. One hundred successes on nominal totes are insufficient. Tests should stratify mass, handle orientation, lighting, floor friction, and aisle traffic, recording success, collision-free abort, automatic recovery, human-intervention minutes, energy per task, and regrasp count. A system that does not fall but repeatedly drops parts has failed. A system that finishes only after remote alignment on each cycle has weak automation economics.

The final step assigns operational ownership. Manufacturing engineering owns task tolerances. Safety owns protective layers and change approval. IT and security own network and update controls. The supplier owns firmware, parts, and repair response. The data team owns logs and model versions. When nobody can fully predict learned behavior, a responsibility matrix defining how to reproduce an incident from exact software, hardware, and sensor state becomes essential.

Manufacturing Cell Checkpoint

The task schema should include target object, start and end pose, permitted contact, forbidden zones, maximum force, timeout, and a safe post-failure posture. “Move a box” is not enough. It should specify which shelf, grasp orientation, placement direction, whether exploratory motion is allowed after loss of visibility, and where the robot retreats when a person enters the aisle.

Logging should connect more than sensor files. Robot serial number, joint, hand, and battery condition, firmware, policy, map, calibration, workpiece lot, and the reason for human intervention belong to one run ID. Simulation reproduction also needs the physical parameters and randomization envelope. Without those links, a team cannot separate policy improvement from a fortunate reduction in hardware variance.

KPIs should measure flows of completed economic work, not peak speed. Include completed units per hour, continuous intervention-free time, protective-stop rate, object damage, automatic recovery, energy per task, consumables, and remote-support cost. Do not inspect only the pilot average. Examine distributions by shift, robot, and workpiece and investigate the worst tail. Scale-up usually fails in tail risk, not in the mean demo.

Safety approval should remain separate from model-performance approval. After each policy update, regression-test that independent stops, speed limits, and workspace monitoring still operate. If a supplier can update remotely, contractually define the approval gate and rollback authority. Recovery after a fall must isolate residual energy and prevent unexpected restart while a person enters the cell.

4.11 Uncertainty and Failure Modes

The first limitation is unequal disclosure. UBTECH's audited figures do not establish technical superiority, and undisclosed private-company revenue does not equal zero. The atlas must accept this asymmetry: separate the existence of a disclosed number from actual performance while refusing to fill absent values with estimates.

The second limitation is attribution between a product family and a research paper. An external paper succeeding on G1 demonstrates research accessibility of the Unitree body; it is not automatically an official product feature or support commitment. AgiBot team simulator results are not independent benchmarks. UBTECH's annual report combines audited statements and issuer narrative, so financial tables, product specifications, and management descriptions carry different evidence grades.

The third limitation is visibility. A complete company-by-company export of eligible articles from the defined media corpus for August 17, 2025 through August 17, 2026 is unavailable. All three visibility fields therefore remain unmeasured. Search-result counts, social-video views, and stage appearances cannot substitute for independently edited article counts. Visibility is also distinct from maturity.

The fourth limitation is partial WRC 2026 verification. The official exhibitor directory verifies Unitree at C215 and UBTECH at C103, while AgiBot's booth and all three program entries remain unverified. A visit plan should still recheck the official directory and on-site signage and should not make an unverified stop mandatory ^[16].

The fifth limitation is safety and long-term reliability. Public demonstrations and short papers rarely cover joint lifetime, hand-pad wear, battery aging, calibration drift, repair-parts availability, or service response. Before a factory purchase, repeat the same task across multiple serial numbers and shifts and request failure videos and service records, not only success footage.

Relation to Prior Surveys

The humanoid locomotion and manipulation literature maps model-based control, learning-based locomotion, whole-body manipulation, and foundation-model research at broad scale ^[14]. This chapter does not project that algorithm map directly onto companies. Moving from research performance to product maturity requires manufacturing variance, documentation, SDKs, independent safety, production, sales, and operating evidence. A strong research body is not necessarily the strongest procurement choice, and the reverse is also true.

Systematic quadruped sim-to-real and humanoid retargeting add two procurement questions before “how good is the policy?” How quickly can the team identify the dynamics of this physical unit? Can it transfer motion from a human or another embodiment while preserving contact relationships? These capabilities determine update cadence and the cost of adding tasks.

What to Learn Next

This chapter treated the general-purpose body as a coupled locomotion-and-manipulation problem. Manufacturing value, however, is ultimately realized in grasping, carrying, and placing. Chapter 5 shifts to mobile manipulation: arms mounted on wheeled or legged bases. A humanoid accepts nearly every difficulty of a human-oriented environment in one body. A mobile manipulator may combine only the mobility and arm capability a task requires, reducing complexity.

Carry three questions into Chapter 5. How does a wheeled, legged, single-arm, or dual-arm form change workspace and stability? Should navigation and grasping remain separate modules or become one whole-body policy? How does a successful demo become intervention-free time and cost per task in a manufacturing cell? The next contest is not a more human-like walk. It is predictable completion of contact work on a moving base.

References

1. AgiBot (2026a). AGIBOT identity and milestones. Official company website.
2. AgiBot (2026b). AgiBot research portfolio. Official research website.
3. Yin, C. et al. (2026). Genie Sim 3.0: A High-Fidelity Comprehensive Simulation Platform for Humanoid Robot. arXiv:2601.02078.
4. Unitree Robotics (2026a). About Unitree. Official company website.
5. Unitree Robotics (2026b). Unitree developer documentation. Official developer website.
6. Unitree Robotics (2026c). Official open-source resources. Official open-source portal.
7. UBTECH Robotics (2026). Annual Report 2025. HKEX issuer filing, stock code 9880.
8. Sferrazza, C. et al. (2024). HumanoidBench: Simulated Humanoid Benchmark for Whole-Body Locomotion and Manipulation. RSS 2024, arXiv:2403.10506.
9. Wu, Z. et al. (2026). Perceptive Humanoid Parkour: Chaining Dynamic Human Skills via Motion Matching. arXiv:2602.15827. #71
10. Seo, Y. et al. (2025). Learning Sim-to-Real Humanoid Locomotion in 15 Minutes. arXiv:2512.01996. #72
11. Bjelonic, F. et al. (2025). Towards Bridging the Gap: Systematic Sim-to-Real Transfer for Diverse Legged Robots. arXiv:2509.06342. #82
12. Yang, L. et al. (2025). OmniRetarget: Interaction-Preserving Data Generation for Humanoid Whole-Body Loco-Manipulation and Scene Interaction. arXiv:2509.26633. #70
13. Cheng, X. et al. (2024). Open-TeleVision: Teleoperation with Immersive Active Visual Feedback. CoRL 2024, arXiv:2407.01512.
14. Gu, Z. et al. (2025). Humanoid Locomotion and Manipulation: Current Progress and Challenges in Control, Planning, and Learning. arXiv:2501.02116. #73
15. Sentis, L. and Khatib, O. (2008). Robot Manipulator Control Using Generalized Operational Space. Robotica. DOI: 10.1017/S0263574708004472.
16. World Robot Conference (2026). World Robot Conference official website. Official organizer source.
17. Jeremy Dao et al. (2022). Sim-to-Real Learning of All Common Bipedal Gaits via Periodic Reward Composition. arXiv:2011.01387.
18. Jonah Siekmann et al. (2021). Blind Bipedal Stair Traversal via Sim-to-Real Reinforcement Learning. arXiv:2105.08328.
19. Nikita Rudin et al. (2021). Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning. arXiv:2109.11978.
20. Jonah Siekmann et al. (2020). Learning Memory-Based Control for Human-Scale Bipedal Locomotion. arXiv:2006.02402.
21. Joonho Lee et al. (2020). Learning Quadrupedal Locomotion Over Challenging Terrain. arXiv:2010.11251.
22. Benjamin Katz et al. (2019). Mini Cheetah: A Platform for Pushing the Limits of Dynamic Quadruped Control. DOI:10.1109/icra.2019.8793865.
23. Jemin Hwangbo et al. (2019). Learning Agile and Dynamic Motor Skills for Legged Robots. DOI:10.1126/scirobotics.aau5872.
24. Jared Di Carlo et al. (2018). Dynamic Locomotion in the MIT Cheetah 3 through Convex Model-Predictive Control. DOI:10.1109/iros.2018.8594448.
25. Patrick M. Wensing et al. (2017). Proprioceptive Actuator Design in the MIT Cheetah: Impact Mitigation and High-Bandwidth Physical Interaction for Dynamic Legged Robots. DOI:10.1109/tro.2016.2640183.
26. Siyuan Feng et al. (2014). Optimization-based Full Body Control for the DARPA Robotics Challenge. DOI:10.1002/rob.21559.
27. Sangok Seok et al. (2013). Design Principles for Highly Efficient Quadrupeds and Implementation on the MIT Cheetah Robot. DOI:10.1109/icra.2013.6631038.
28. Ludovic Righetti et al. (2013). Control of Legged Robots with Optimal Distribution of Contact Forces. DOI:10.1177/0278364913495243.
29. Jerry Pratt et al. (2012). Capturability-based Analysis and Control of Legged Locomotion, Part 2: Application to M2V2, a Lower-Body Humanoid. DOI:10.1177/0278364912452762.
30. Twan Koolen et al. (2012). Capturability-based Analysis and Control of Legged Locomotion, Part 1: Theory and Application to Three Simple Gait Models. DOI:10.1177/0278364912452673.
31. Eric R. Westervelt et al. (2007). Feedback Control of Dynamic Bipedal Robot Locomotion. CRC Press (Book).
32. Jerry Pratt et al. (2006). Capture Point: A Step toward Humanoid Push Recovery. DOI:10.1109/ichr.2006.321385.
33. Luis Sentis et al. (2005). Synthesis of Whole-Body Behaviors through Hierarchical Control of Behavioral Primitives. DOI:10.1142/s0219843605000545.
34. Shuuji Kajita et al. (2003). Biped walking pattern generation by using preview control of zero-moment point. DOI:10.1109/robot.2003.1241826.
35. David W. Robinson (2000). Design and Analysis of Series Elasticity in Closed-loop Actuator Force Control. MIT PhD thesis.
36. Gill A. Pratt et al. (1995). Series Elastic Actuators. DOI:10.1109/iros.1995.525827.