Pith. sign in

REVIEW 3 major objections 5 minor 172 references

This review claims that legged locomotion — especially quadrupedal walking — has matured to the point where the next bottleneck is semantic understanding and dexterous interaction, not basic walking.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 16:35 UTC pith:IWKD5E4T

load-bearing objection A competent, well-written review whose useful 'dexterous semantic locomotion' framing is undercut by a conclusion that overstates locomotion reliability relative to its own Figure 2B and data section. the 3 major comments →

arxiv 2607.28952 v1 pith:IWKD5E4T submitted 2026-07-31 cs.RO

Advances, challenges, and opportunities for legged robots

Jonas Frey , Mat\'ias Mattamala , Hae-Won Park , Mayank Mittal , Georg Martius , Maike Osborne , Robert Sparrow , Marco Hutter This is my paper
classification cs.RO
keywords legged robotsquadrupedal locomotionhumanoid robotsreinforcement learningdexterous semantic locomotionsim-to-realterrain understandingrobot policy
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The review argues that legged robot hardware and reinforcement-learning-based control have matured enough to make quadrupedal walking an approachable problem, and that the field's frontier has shifted to semantic understanding and dexterous interaction: reading terrain not just geometrically but semantically, and placing feet precisely in response. It surveys hardware, locomotion, autonomy, data, applications, ethics, and economics to support this shift. A sympathetic reader should care because if true, the near-term agenda for research and industry should concentrate on semantic terrain understanding, precise foot placement, and cross-domain dexterity rather than on basic walking. The paper also argues that legged robots, by operating in human environments and service sectors, could affect far more of the workforce than industrial robots did, and it recommends capability-based regulation.

Core claim

The paper's central claim is that legged locomotion has matured to broad applicability: electromagnetic actuation with low gear ratios and backdrivable torque control, combined with simulation-trained reinforcement learning policies, has made quadrupedal walking reliable across diverse terrains. It therefore proposes that the field's frontier has shifted to what it names 'dexterous semantic locomotion' — the ability to interpret the environment semantically (anticipating loose stones, branches, social conventions, multi-agent interactions) and to respond with precise, environment-conditioned foot placement and dexterous interaction, rather than merely reacting to geometry. The review contend

What carries the argument

The paper's central organizing concept is 'dexterous semantic locomotion,' a term it coins to name the emerging paradigm in which legged robots anticipate interactions with terrain beyond geometry and interpret multimodal cues to plan fine-grained motor responses. The other load-bearing mechanism is the sim-to-real reinforcement learning pipeline: policies trained in physics simulators with domain randomization and privileged information, then transferred zero-shot, which is what elevated locomotion from hand-engineered control to robust, learnable behavior. The review uses these two mechanisms to structure its argument that hardware solved the actuation problem, reinforcement learning solve

Load-bearing premise

The load-bearing premise is that today's real-world capability is accurately reflected in vendor press releases and selected demonstrations, rather than in independently benchmarked deployments.

What would settle it

Run a standardized benchmark that compares foot-placement success on terrain with identical geometry but different semantic labels (for example, loose gravel versus solid rock, or a glass panel versus concrete): if success rates do not differ once geometry is equated, the claim that semantic understanding is the next frontier would be undermined; if they do, the claim is supported.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Quadrupedal robots are commercially viable for inspection, delivery, and security today, and falling hardware prices will push them into new markets; the same platforms will increasingly be judged on semantic understanding and precise interaction rather than raw mobility.
  • Humanoid development should de-emphasize purely reactive locomotion and focus on precise foot placement and nuanced environment understanding, which are prerequisites for whole-body manipulation and home use.
  • Sim-to-real training must expand beyond rigid-body simulation, because current simulators cannot model entanglement in vegetation, deformable ground, or the correct physical interaction behind visual semantics.
  • Autonomy architecture will shift from modular, hand-crafted interfaces toward fused, learned representations with language and vision priors, while real-time control-frequency constraints remain.
  • Policy should be capability-based: minimal oversight for walking-only robots, escalating with manipulation and social interaction; international coordination is needed to manage military and lethal legged-robot risks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If semantic understanding is truly the bottleneck, a testable corollary is that foot-placement failures on geometrically identical but semantically different terrain should dominate over dynamics failures; publishing such failure decompositions would sharpen or refute the review's thesis.
  • The paper's market outlook implicitly assumes that current commercial pilots (delivery, home humanoids) scale without the safety and verification gap it acknowledges; a more cautious reading is that near-term economics rest on teleoperation and shared autonomy rather than full autonomy.
  • The 'dexterous semantic locomotion' frame could be operationalized as a benchmark: measure success on terrain where geometry and semantics disagree (for example, frozen lake versus solid ground, or glass panel versus concrete), which would connect legged locomotion to established semantic-segmentation research.
  • The review's data-flywheel argument implies that teleoperated deployment of legged robots could generate embodied datasets at internet scale, but only if the community adopts shared data formats and benchmarks; the paper notes the lack of standardized benchmarks but does not propose one.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. This review synthesizes the state of legged robotics across hardware, locomotion, autonomy, data, applications, ethics, and policy. It argues that advances in actuation, RL-based sim-to-real control, and sensing have brought legged robots to the threshold of broad deployment, and that the field's frontier has shifted from basic locomotion to 'dexterous semantic locomotion'—combining semantic terrain understanding with precise foot placement and manipulation-like interaction. It closes with recommendations for capability-based regulation, industrial policy, and workforce programs.

Significance. The review is timely and potentially agenda-setting. Its strengths are the breadth of coverage; the clear articulation of the 'dexterous semantic locomotion' paradigm; the data pyramid/flywheel framing; the integration of technical, ethical, and economic dimensions; and an honest treatment of open problems such as the sim-to-real visual gap, deformable terrain, and lack of standardized benchmarks. The main risk is that the conclusion's maturity claim is stronger than the evidence assembled in the body of the paper. If the authors calibrate that claim and qualify vendor-based capability evidence, the result would be a balanced and useful reference.

major comments (3)
  1. [Conclusion vs. Fig. 2B and Data and Simulations] The central claim that 'RL has made quadrupedal locomotion an approachable problem, yielding reliable performance across diverse environments' (Conclusion) is not supported by the paper's own evidence. Fig. 2B's caption states that 'current robots reliably navigate flat ground and increasingly traverse rough and discontinuous terrains,' and the Data and Simulations section lists deformable terrain, entanglement in vegetation, and semantically rich cluttered environments as beyond current simulators, with an unresolved sim-to-real visual gap. 'Increasingly' is not 'reliably.' Because this overstatement is the basis for declaring the frontier shifted to semantic understanding and dexterity, the conclusion should be revised to state that locomotion is reliable primarily on flat or geometrically known terrain and that robust operation on natural unstructured terrain remains an open problem.
  2. [Introduction and Real-World Applications, refs 1, 2, 138] Several load-bearing capability and application claims rest on vendor press releases and public demonstrations rather than independent evaluation. The Introduction states that legged robots 'deliver parcels directly to front doors (1,2)' and the Real-World Applications section forecasts 'commercial pilot humanoids... by 2026 (138)'; refs 1, 2, and 138 are respectively a company press page, a vendor demonstration, and a pre-order page. These sources may be accurate, but a Science Robotics review should attribute such claims as vendor statements or support them with peer-reviewed deployments (e.g., refs 131, 133, 137). Without this, the optimistic capability assessment is vulnerable to overstatement.
  3. [Policy and Economics (price ranges and capability timeline)] Quantitative economic claims are not systematically sourced. The $30,000–$90,000 price band for current quadrupeds and the $2,700/$4,900 entry-level prices rest on refs 37 and 166, which are vendor media items; the 'plausible 10- to 15-year progression from robot walking to robot social capabilities' is presented without a citation or derivation. The 47% automation-risk and 0.42% wage-effect statistics (refs 167, 168) are cited correctly but are transported from general computerization/manufacturing contexts to legged robots without discussion of external validity. For a policy-oriented review, these numbers should be labeled as illustrative and either sourced to independent market analyses or removed.
minor comments (5)
  1. [Header/title page] The header 'ReviewReview Paper' is duplicated on the title page and running head.
  2. [Fig. 1C] 'Historic development' should be 'Historical development.'
  3. [Fig. 2B and Conclusion] The caption's 'reliably... flat ground' and 'increasingly traverse rough and discontinuous terrains' should be mirrored in the Conclusion; see major comment. Also define 'perceptive vision' on the x-axis.
  4. [Real-World Applications, ref 102] The claim that the LS3/BigDog period 'involved the largest field deployments ever reported for a legged platform (102)' is not supported by ref 102, which is a scene-understanding paper; either cite the appropriate field-deployment report or soften the claim.
  5. [Data and Simulations, ref 122] The text refers to the 'GrandTour Dataset' but ref 122 is the 'Boxi' dataset paper. Please correct the citation or the dataset name.

Circularity Check

0 steps flagged

No circularity: review synthesizes cited evidence; the only notable tension is an internal over-claim, not a circular derivation.

full rationale

This is a review article with no equations, fitted parameters, numerical predictions, or formal derivation chains. Its load-bearing claims—that RL has made quadrupedal locomotion reliable, that the frontier is shifting toward semantic understanding and dexterity, and that delivery/inspection applications are maturing—are supported by citations to peer-reviewed, independently checkable work (e.g., refs 11, 43, 62, 110, 122, 123) and by vendor demonstrations. Self-citations from the Hutter group are frequent, but they are used as ordinary literature evidence for claims with external reproducibility (videos, code, field deployments), not as a uniqueness theorem or as a postulated ansatz that forecloses alternatives. The coined term 'dexterous semantic locomotion' (Locomotion section, p. 5) is explicitly presented as a label for an emerging research paradigm, not as a new empirical result derived from first principles; naming a proposed direction is not circular. The closest thing to a concern is internal inconsistency rather than circularity: the Conclusion states that RL yields 'reliable performance across diverse environments,' while Figure 2B's caption restricts reliability to flat ground and says rough and discontinuous terrains are only 'increasingly' traversed; the Data and Simulations section also lists deformable terrain, vegetation entanglement, and semantically rich cluttered environments as beyond current simulators. That is an over-claim or correctness issue, not a reduction of a prediction to its inputs. No circular step can be exhibited, so the circularity score is 0.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The review does not fit parameters or introduce mathematical assumptions. Its central narrative rests on the accuracy of the cited literature and vendor-reported deployments, plus a projected capability timeline for policy recommendations. No invented physical entities are introduced.

axioms (2)
  • domain assumption The cited peer-reviewed and vendor-reported sources accurately represent the state of the art in legged robotics.
    The review's capability assessment in Hardware, Locomotion, Data, and Real-World Applications is a synthesis of these sources; no independent verification is provided.
  • domain assumption A 10-to-15-year progression from walking to manipulation to social capabilities will occur at the assumed pace.
    The policy recommendations in Policy and Economics rest on this compressed timeline, described as 'the plausible 10- to 15-year progression from robot walking to robot social capabilities'.

pith-pipeline@v1.3.0-daily-deepseek · 30839 in / 9445 out tokens · 92391 ms · 2026-08-03T16:35:02.492342+00:00 · methodology

0 comments
read the original abstract

Humanoid and quadrupedal robots have the potential to revolutionize the way we work, interact, and coexist with intelligent machines. To understand their effects on society and how they can enable scientific discovery, we assess the current capabilities of these systems along hardware, locomotion, autonomy, data, and applications. We identify recent advances and key open challenges that must be overcome to enable widespread adoption and new use cases for legged robots. Last, we provide an outlook on the future of legged robots, exploring their ethical considerations, economic potential, policy implications, and broader societal effects.

Figures

Figures reproduced from arXiv: 2607.28952 by Georg Martius, Hae-Won Park, Jonas Frey, Maike Osborne, Marco Hutter, Mat\'ias Mattamala, Mayank Mittal, Robert Sparrow.

Figure 1
Figure 1. Figure 1: Hardware overview. (A) A comparison of the functional analogies between human body parts (for example, eye and muscles) and robot components (perception and actuation). (B) The internal structure of a robotic actuation unit (18). (C) Historic development of the control paradigm toward RL. (D) The growing research interest in humanoid and quadrupedal robots (19,20,21). (E) Payload capacity versus weight of … view at source ↗
Figure 2
Figure 2. Figure 2: Overview of legged locomotion. (A) The two main phases of the sim-to-real paradigm. During training, a policy is optimized via RL in a simulator to output actuator position targets. A motor controller typically converts these targets into the required torques, which are provided to the rigid body–physics simulator to compute the next state, observation, and reward. Through the use of domain randomization a… view at source ↗
Figure 3
Figure 3. Figure 3: Evolution of autonomy systems. Classical autonomy systems decompose functionality into distinct submodules and process sensory information hierarchically, using clearly defined interfaces (illustrated as gray bounding boxes). Learning-based autonomy systems, by contrast, introduce tightly coupled or fused modules, illustrated in blue. Large language models (LLMs), visual-language models (VLMs), and other j… view at source ↗
Figure 4
Figure 4. Figure 4: Data and simulation overview. (A) The data scaling pyramid illustrates that data exist at varying levels of scale. The more task- and embodiment-specific the data are, the more costly they become to collect. Different data sources support learning different underlying concepts, which can be categorized as world knowledge, general knowledge, and task-specific knowledge. (B) A modern data flywheel demonstrat… view at source ↗
Figure 5
Figure 5. Figure 5: Applications of legged robots. (A) Digit robot performs last-mile delivery, and Electric Atlas by Boston Dynamics performs mobile manipulation for an industrial task. (B) Boston Dynamics’ Spot monitors a construction site, and ANYmal D performs industrial inspection at a manufacturing plant. (C) Boston Dynamics LS3 acting as a mule in rough terrain, Unitree B2 tested by a fire brigade for situational aware… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

172 extracted references · 27 canonical work pages · 1 internal anchor

  1. [1]

    ai/, press Release

    RIVR, RIVR: Pioneering Physical AI for Last-Mile De- livery, https://www.rivr.ai/ (2025), https://www.rivr. ai/, press Release. Accessed: 2025-10

    2025

  2. [2]

    com/(2025), public Demonstration

    Agility Robotics, Digit, https://www.agilityrobotics. com/(2025), public Demonstration. Accessed: 2025-11

    2025

  3. [3]

    Tranzatto, T

    M. Tranzatto, T. Miki, M. Dharmadhikari, L. Bern- reiter, M. Kulkarni, F. Mascarich, O. Andersson, S. Khattak, M. Hutter, R. Siegwart, K. Alexis, CERBERUS in the DARPA Subterranean Challenge. Sci. Robot.7(66), eabp9742 (2022), doi:10.1126/ scirobotics.abp9742, https://www.science.org/doi/abs/ 10.1126/scirobotics.abp9742

    doi:10.1126/scirobotics.abp9742 2022

  4. [4]

    M. H. Raibert,Legged Robots That Balance(MIT Press, Cambridge, MA) (1985), doi:10.1109/MEX.1986.4307016

    arXiv 1985

  5. [5]

    Todd,Walking Machines: An Introduction to Legged Robots, Advanced Industrial Technology Series (Chapman and Hall) (1985), doi:https://doi.org/10.1007/ 978-1-4684-6858-8

    D. Todd,Walking Machines: An Introduction to Legged Robots, Advanced Industrial Technology Series (Chapman and Hall) (1985), doi:https://doi.org/10.1007/ 978-1-4684-6858-8

    1985

  6. [6]

    Crowley, J

    D. Crowley, J. Dao, H. Duan, K. Green, J. Hurst, A. Fern, Optimizing Bipedal Locomotion for The 100m Dash With Comparison to Human Running, in2023 IEEE International Conference on Robotics and Automation (ICRA)(2023), pp. 12205–12211, doi:10.1109/ICRA48891.2023.10160436

    arXiv 2023

  7. [7]

    Sakagami, R

    Y. Sakagami, R. Watanabe, C. Aoyama, S. Matsunaga, N. Higaki, K. Fujimura, The intelligent ASIMO: system overview and integration, inIEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS)(2002), pp. 2478– 2483, doi:10.1109/IRDS.2002.1041641, https://doi.org/ 10.1109/IRDS.2002.1041641. 12 preprint for Science Robotics | ReviewReview Paper 2026

    Pith/arXiv arXiv 2002

  8. [8]

    Raibert, K

    M. Raibert, K. Blankespoor, G. Nelson, R. Playter, BigDog, the Rough-Terrain Quadruped Robot.IFAC Proceedings Volumes41(2), 10822–10825 (2008), 17th IFAC World Congress, doi:https://doi.org/10.3182/ 20080706-5-KR-1001.01833, https://www.sciencedirect. com/science/article/pii/S1474667016407020

    arXiv 2008

  9. [9]

    B. G. Katz,A Low Cost Modular Actuator for Dynamic Robots, S.M. Thesis, Massachusetts Institute of Technology, Department of Mechanical Engineering (2018), electronic version submitted by the student author. Available at:http: //dspace.mit.edu/handle/1721.1/7582

    2018

  10. [10]

    Grimminger, A

    F. Grimminger, A. Meduri, M. Khadiv, J. Viereck, M. Wüthrich, M. Naveau, V. Berenz, S. Heim, F. Wid- maier, T. Flayols, J. Fiene, A. Badri-Spröwitz, L. Righetti, An Open Torque-Controlled Modular Robot Architecture for Legged Locomotion Research.IEEE Robot. Autom. Lett. (RA-L)5(2), 3650–3657 (2020), doi:10.1109/LRA.2020. 2976639

    doi:10.1109/lra.2020 2020

  11. [11]

    Hwangbo, J

    J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V. Tsou- nis, V. Koltun, M. Hutter, Learning agile and dynamic motor skills for legged robots.Sci. Robot.4(26), eaau5872 (2019), doi:10.1126/scirobotics.aau5872, https://www. science.org/doi/abs/10.1126/scirobotics.aau5872

    doi:10.1126/scirobotics.aau5872 2019

  12. [12]

    J. Tan, T. Zhang, E. Coumans, A. Iscen, Y. Bai, D. Hafner, S. Bohez, V. Vanhoucke, Sim-to-Real: Learning Agile Lo- comotion For Quadruped Robots, inRobotics: Science and Systems (RSS), H. Kress-Gazit, S. S. Srinivasa, T. Howard, N. Atanasov, Eds. (2018), doi:10.15607/RSS.2018.XIV.010, http://www.roboticsproceedings.org/rss14/p10.html

    doi:10.15607/rss.2018.xiv.010 2018

  13. [13]

    Todorov, T

    E. Todorov, T. Erez, Y. Tassa, MuJoCo: A physics engine for model-based control, inIEEE/RSJ Intl. Conf. on Intel- ligent Robots and Systems (IROS)(2012), pp. 5026–5033, doi:10.1109/IROS.2012.6386109

    arXiv 2012

  14. [14]

    Makoviychuk, L

    V. Makoviychuk, L. Wawrzyniak, Y. Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, G. State, Isaac Gym: High Performance GPU Based Physics Simulation For Robot Learning, inIntl. Conf. on Neural Information Processing Systems (NeurIPS), J. Vanschoren, S. Yeung, Eds. (2021)

    2021

  15. [15]

    Rudin, D

    N. Rudin, D. Hoeller, P. Reist, M. Hutter, Learning to Walk in Minutes Using Massively Parallel Deep Rein- forcement Learning, inConf. on Robot Learning (CoRL), A. Faust, D. Hsu, G. Neumann, Eds., vol. 164 ofPro- ceedings of Machine Learning Research(2021), pp. 91–100, https://proceedings.mlr.press/v164/rudin22a.html

    2021

  16. [16]

    Mittal, P

    M. Mittal, P. Roth, J. Tigue, A. Richard, O. Zhang, P. Du, A. Serrano-Muñoz, X. Yao, R. Zurbrügg, N. Rudin, L. Wawrzyniak, M. Rakhsha, A. Denzler, E. Heiden, A. Borovicka, O. Ahmed, I. Akinola, A. Anwar, M. T. Carlson, J. Y. Feng, A. Garg, R. Gasoto, L. Gulich, Y. Guo, M. Gussert, A. Hansen, M. Kulkarni, C. Li, W. Liu, V. Makoviychuk, G. Malczyk, H. Mazha...

    Pith/arXiv arXiv 2025

  17. [17]

    Zakka, B

    K. Zakka, B. Tabanpour, Q. Liao, M. Haiderbhai, S. Holt, J. Y. Luo, A. Allshire, E. Frey, K. Sreenath, L. A. Kahrs, C. Sferrazza, Y. Tassa, P. Abbeel, Demonstrating Mu- JoCo Playground, inRobotics: Science and Systems (RSS) (LosAngeles, CA, USA) (2025), doi:10.15607/RSS.2025.XXI. 020

    doi:10.15607/rss.2025.xxi 2025

  18. [19]

    humanoid OR biped OR bipedal

    OpenAlex, Works filtered by title and abstract: "humanoid OR biped OR bipedal", publication years 1980–2025, https://openalex.org/works?page=1&filter=title_ and_abstract.search:humanoid+OR+biped+OR+bipedal, publication_year:1980-2025(2025), accessed: 2025-11

    1980

  19. [20]

    quadruped OR quadrupedal

    OpenAlex, Works filtered by title and abstract: "quadruped OR quadrupedal", publication years 1980–2025, https://openalex.org/works?page=1&filter=title_ and_abstract.search:quadruped+OR+quadrupedal, publication_year:1980-2025(2025), accessed: 2025-11

    1980

  20. [21]

    hexapod OR octopod OR multi-legged

    OpenAlex, Works filtered by title and abstract: "hexapod OR octopod OR multi-legged", publication years 1980– 2025, https://openalex.org/works?page=1&filter= title_and_abstract.search:hexapod+OR+octopod+or+ multi-legged,publication_year:1980-2025 (2025), accessed: 2025-11

    1980

  21. [22]

    H. Park, P. M. Wensing, S. Kim, High-speed bound- ing with the MIT Cheetah 2: Control design and experiments.Intl. J. of Robot. Res.36(2), 167–192 (2017), doi:10.1177/0278364917694244, https://doi.org/ 10.1177/0278364917694244

    doi:10.1177/0278364917694244 2017

  22. [23]

    P. M. Wensing, A. Wang, S. Seok, D. Otten, J. H. Lang, S. Kim, Proprioceptive Actuator Design in the MIT Cheetah: Impact Mitigation and High-Bandwidth Physical Interaction for Dynamic Legged Robots.IEEE Trans. Robot.33(3), 509–522 (2017), doi:10.1109/TRO.2016.2640183,https:// doi.org/10.1109/TRO.2016.2640183

    arXiv 2017

  23. [24]

    Siciliano, O

    B. Siciliano, O. Khatib, eds.,Springer Handbook of Robotics, Springer Handbooks (Springer) (2016), doi: 10.1007/978-3-319-32552-1, https://doi.org/10.1007/ 978-3-319-32552-1

    doi:10.1007/978-3-319-32552-1 2016

  24. [25]

    G. A. Pratt, M. M. Williamson, Series elastic actuators, in IEEE/RSJ Intl. Conf. on Intelligent Robots and Systems (IROS), vol. 1 (1995), pp. 399–406

    1995

  25. [26]

    J. M. Hollerbach, I. W. Hunter, J. Ballantyne,A comparative analysis of actuator technologies for robotics(MIT Press, Cambridge, MA, USA), pp. 299–342 (1992)

    1992

  26. [27]

    G. A. Pratt, M. M. Williamson, P. Dillworth, J. Pratt, A. Wright, Stiffness isn’t everything, inExperimental Robotics IV, O. Khatib, J. K. Salisbury, Eds. (Springer Berlin Heidelberg, Berlin, Heidelberg) (1997), pp. 253–262. 13 preprint for Science Robotics | ReviewReview Paper 2026

    1997

  27. [28]

    Englsberger, A

    J. Englsberger, A. Werner, C. Ott, B. Henze, M. A. Roa, G. Garofalo, R. Burger, A. Beyer, O. Eiberger, K. Schmid, A. Albu-Schäffer, Overview of the torque-controlled hu- manoid robot TORO, inIEEE-RAS International Con- ference on Humanoid Robots(2014), pp. 916–923, doi: 10.1109/HUMANOIDS.2014.7041473

    arXiv 2014

  28. [29]

    Pratt, B

    J. Pratt, B. Krupp, Design of a bipedal walking robot. Proceedings of SPIE - The International Society for Optical Engineering6962(2008), doi:10.1117/12.777973

    work page doi:10.1117/12.777973 2008

  29. [30]

    Hutter,StarlETH & Co

    M. Hutter,StarlETH & Co. – Design and Control of Legged Robots with Compliant Actuation, Ph.D. thesis, ETH Zurich, Zurich, Switzerland (2013)

    2013

  30. [31]

    R. W. David,Design and analysis of series elastic- ity in closed-loop actuator force control, Ph.D. thesis, Massachusetts Institute of Technology, Cambridge, Mas- sachusetts, United States (2000)

    2000

  31. [32]

    M. H. Raibert, Trotting, pacing and bounding by a quadruped robot.Journal of Biomechanics23, 79–98 (1990), international Society of Biomechan- ics, doi:https://doi.org/10.1016/0021-9290(90)90043-3, https://www.sciencedirect.com/science/article/pii/ 0021929090900433

    work page doi:10.1016/0021-9290(90)90043-3 1990

  32. [33]

    Accessed: 2025-10

    Boston Dynamics, Atlas, The Next Generation, https: //www.youtube.com/watch?v=rVlhMGQgDkY (2016), public Demonstration. Accessed: 2025-10

    2016

  33. [34]

    Semini, N

    C. Semini, N. G. Tsagarakis, E. Guglielmino, M. Focchi, F. Cannella, D. G. Caldwell, Design of HyQ–a hydraulically and electrically actuated quadruped robot.Proceedings of the Institution of Mechanical Engineers, Part I: Journal of Systems and Control Engineering225(6), 831–849 (2011), doi:10.1177/0959651811402275

    doi:10.1177/0959651811402275 2011

  34. [35]

    Hutter, C

    M. Hutter, C. Gehring, D. Jud, A. Lauber, C. D. Belli- coso, V. Tsounis, J. Hwangbo, K. Bodie, P. Fankhauser, M. Bloesch, R. Diethelm, S. Bachmann, A. Melzer, M. Hoepflinger, ANYmal - a highly mobile and dynamic quadrupedal robot, inIEEE/RSJ Intl. Conf. on Intelli- gent Robots and Systems (IROS)(2016), pp. 38–44, doi: 10.1109/IROS.2016.7758092

    arXiv 2016

  35. [36]

    S. Seok, A. Wang, D. Otten, S. Kim, Actuator design for high force proprioceptive control in fast legged locomotion, inIEEE/RSJ Intl. Conf. on Intelligent Robots and Sys- tems (IROS)(2012), pp. 1970–1975, doi:10.1109/IROS.2012. 6386252

    doi:10.1109/iros.2012 2012

  36. [37]

    youtube.com/watch?v=emF9BW0DHoc (2020), press Release

    Unitree Robotics, The Preview of Unitree A1,https://www. youtube.com/watch?v=emF9BW0DHoc (2020), press Release. Accessed: 2025-10

    2020

  37. [38]

    Accessed: 2025-10

    Unitree Robotics, Unitree Introducing | Unitree G1 Hu- manoid Agent | AI Avatar | Price from $16K, https: //www.youtube.com/watch?v=GzX1qOIO1bE (2024), public Demonstration. Accessed: 2025-10

    2024

  38. [39]

    T. Zhu, M. S. Ahn, D. Hong, ARTEMIS: An Open-Source, Full-Sized Humanoid Robot for Dynamic Locomotion, in IEEE-RAS International Conference on Humanoid Robots (2025), pp. 269–276, doi:10.1109/Humanoids65713.2025. 11203020

    arXiv 2025

  39. [40]

    Bloesch, M

    M. Bloesch, M. Hutter, M. Hoepflinger, S. Leutenegger, C. Gehring, C. D. Remy, R. Siegwart, State Estimation for Legged Robots - Consistent Fusion of Leg Kinematics and IMU, inRobotics: Science and Systems (RSS)(Sydney, Australia) (2012), doi:10.15607/RSS.2012.VIII.003

    doi:10.15607/rss.2012.viii.003 2012

  40. [41]

    Fankhauser, M

    P. Fankhauser, M. Bloesch, M. Hutter, Probabilistic Terrain Mapping for Mobile Robots With Uncertain Localization. IEEE Robot. Autom. Lett. (RA-L)3(4), 3019–3026 (2018), doi:10.1109/LRA.2018.2849506

    arXiv 2018

  41. [42]

    Accessed: 2025-10

    Boston Dynamics, Atlas | Partners in Parkour, https: //youtu.be/tF4DML7FIWk (2021), public Demonstration. Accessed: 2025-10

    2021

  42. [43]

    T. Miki, J. Lee, J. Hwangbo, L. Wellhausen, V. Koltun, M. Hutter, Learning robust perceptive locomotion for quadrupedal robots in the wild.Sci. Robot.7(62), eabk2822 (2022), doi:10.1126/scirobotics.abk2822, https://www. science.org/doi/abs/10.1126/scirobotics.abk2822

    doi:10.1126/scirobotics.abk2822 2022

  43. [44]

    Kumar, Z

    A. Kumar, Z. Fu, D. Pathak, J. Malik, RMA: Rapid Motor Adaptation for Legged Robots, inRobotics: Science and Systems (RSS)(Virtual) (2021), doi:10.15607/RSS.2021. XVII.011

    doi:10.15607/rss.2021 2021

  44. [45]

    H.-W. Park, K. Sreenath, J. W. Hurst, J. W. Grizzle, Iden- tification of a Bipedal Robot with a Compliant Drivetrain. IEEE Control Systems Magazine31(2), 63–88 (2011), doi: 10.1109/MCS.2010.939963

    arXiv 2011

  45. [46]

    T. Jung, J. Lim, H. Bae, K. K. Lee, H.-M. Joe, J.-H. Oh, Development of the Humanoid Disaster Response Platform DRC-HUBO+.IEEE Trans. Robot.34(1), 1–17 (2018), doi:10.1109/TRO.2017.2776287

    arXiv 2018

  46. [47]

    H.-W. Park, A. Ramezani, J. W. Grizzle, A Finite-State Ma- chine for Accommodating Unexpected Large Ground-Height Variations in Bipedal Robot Walking.IEEE Trans. Robot. 29(2), 331–345 (2013), doi:10.1109/TRO.2012.2230992

    arXiv 2013

  47. [48]

    F. Liu, S. Deswal, A. Christou, Y. Sandamirskaya, M. Kaboli, R. Dahiya, Neuro-inspired electronic skin for robots.Sci. Robot.7(67), eabl7344 (2022), doi:10.1126/ scirobotics.abl7344, https://www.science.org/doi/abs/ 10.1126/scirobotics.abl7344

    work page doi:10.1126/scirobotics.abl7344 2022

  48. [49]

    T. J. K. Buchner, T. Fukushima, A. Kazemipour, S.-D. Gravert, M. Prairie, P. Romanescu, P. Arm, Y. Zhang, X. Wang, S. L. Zhang, J. Walter, C. Keplinger, R. K. Katzschmann, Electrohydraulic musculoskeletal robotic leg for agile, adaptive, yet energy-efficient locomotion.Na- ture Communications15(1), 7634 (2024), doi:10.1038/ s41467-024-51568-3

    2024

  49. [50]

    Z. Hu, Y. Zhang, H. Jiang, J. an Lv, Bioinspired helical- artificial fibrous muscle structured tubular soft actua- tors.Science Advances9(25), eadh3350 (2023), doi: 10.1126/sciadv.adh3350, https://www.science.org/doi/ abs/10.1126/sciadv.adh3350

    work page doi:10.1126/sciadv.adh3350 2023

  50. [51]

    Asano, K

    Y. Asano, K. Okada, M. Inaba, Design principles of a human mimetic humanoid: Humanoid platform to study human intelligence and internal body sys- tem.Sci. Robot.2(13), eaaq0899 (2017), doi:10.1126/ scirobotics.aaq0899, https://www.science.org/doi/abs/ 10.1126/scirobotics.aaq0899

    work page doi:10.1126/scirobotics.aaq0899 2017

  51. [52]

    O. Yasa, Y. Toshimitsu, M. Y. Michelis, L. S. Jones, M. Filippi, T. Buchner, R. K. Katzschmann, An Overview of Soft Robotics.Annu. Rev. Control. Robotics Auton. Syst.6(Volume 6, 2023), 1–29 (2023), doi: https://doi.org/10.1146/annurev-control-062322-100607, https://www.annualreviews.org/content/journals/10. 1146/annurev-control-062322-100607

    doi:10.1146/annurev-control-062322-100607 2023

  52. [53]

    Siegwart, I

    R. Siegwart, I. R. Nourbakhsh, D. Scaramuzza,Introduction to Autonomous Mobile Robots, Second Edition, Intelligent robotics and autonomous agents (MIT Press) (2011). 14 preprint for Science Robotics | ReviewReview Paper 2026

    2011

  53. [54]

    Holmes, R

    P. Holmes, R. J. Full, D. Koditschek, J. Guckenheimer, The Dynamics of Legged Locomotion: Models, Analyses, and Challenges.SIAM Review48(2), 207–304 (2006), doi: 10.1137/S0036144504445133, https://doi.org/10.1137/ S0036144504445133

    work page doi:10.1137/s0036144504445133 2006

  54. [55]

    Kashiri, A

    N. Kashiri, A. Abate, S. J. Abram, A. Albu-Schaffer, P. J. Clary, M. Daley, S. Faraji, R. Furnemont, M. Garabini, H. Geyer, A. M. Grabowski, J. Hurst, J. Malzahn, G. Mathijssen, D. Remy, W. Roozing, M. Shahbazi, S. N. Simha, J.-B. Song, N. Smit-Anseeuw, S. Strami- gioli, B. Vanderborght, Y. Yesilevskiy, N. Tsagarakis, An Overview on Principles for Energy ...

    arXiv 2018

  55. [56]

    Reher, A

    J. Reher, A. D. Ames, Dynamic Walking: Toward Agile and Efficient Bipedal Robots.Annu. Rev. Control. Robotics Auton. Syst.4(Volume 4, 2021), 535–572 (2021), doi: https://doi.org/10.1146/annurev-control-071020-045021, https://www.annualreviews.org/content/journals/10. 1146/annurev-control-071020-045021

    work page doi:10.1146/annurev-control-071020-045021 2021

  56. [57]

    Semini, V

    C. Semini, V. Barasuol, T. Boaventura, M. Frigerio, M. Foc- chi, D. G. Caldwell, J. Buchli, Towards versatile legged robots through active impedance control.The International Journal of Robotics Research34(7), 1003–1020 (2015)

    2015

  57. [58]

    Jenelten, J

    F. Jenelten, J. He, F. Farshidian, M. Hutter, DTC: Deep Tracking Control.Sci. Robot.9(86) (2024), doi:10. 1126/scirobotics.adh5401, http://dx.doi.org/10.1126/ scirobotics.adh5401

    2024

  58. [59]

    S. Ha, J. Lee, M. van de Panne, Z. Xie, W. Yu, M. Khadiv, Learning-based legged locomotion: State of the art and future perspectives.Intl. J. of Robot. Res.44(8), 1396– 1427 (2025)

    2025

  59. [60]

    Huang, J

    T. Huang, J. Ren, H. Wang, Z. Wang, Q. Ben, M. Wen, X. Chen, J. Li, J. Pang, Learning Humanoid Standing-up Control across Diverse Postures, inProceedings of Robotics: Science and Systems(LosAngeles, CA, USA) (2025), doi: 10.15607/RSS.2025.XXI.064

    work page doi:10.15607/rss.2025.xxi.064 2025

  60. [61]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, O. Klimov, Proximal Policy Optimization Algorithms (2017),https: //arxiv.org/abs/1707.06347

    Pith/arXiv arXiv 2017

  61. [62]

    J. Lee, J. Hwangbo, L. Wellhausen, V. Koltun, M. Hut- ter, Learning quadrupedal locomotion over challenging ter- rain.Sci. Robot.5(47), eabc5986 (2020), doi:10.1126/ scirobotics.abc5986, https://www.science.org/doi/abs/ 10.1126/scirobotics.abc5986

    doi:10.1126/scirobotics.abc5986 2020

  62. [63]

    Huang, Y

    X. Huang, Y. Chi, R. Wang, Z. Li, X. B. Peng, Y. S. Shao, B. Nikolic, K. Sreenath, DiffuseLoco: Real-Time Legged Locomotion Control with Diffusion from Offline Datasets, inConf. on Robot Learning (CoRL), P. Agrawal, O. Kroemer, W. Burgard, Eds., vol. 270 ofProceedings of Machine Learning Research(2024), pp. 1567–1589,https: //proceedings.mlr.press/v270/hu...

    2024

  63. [64]

    Radosavovic, T

    I. Radosavovic, T. Xiao, B. Zhang, T. Darrell, J. Malik, K. Sreenath, Real-world humanoid locomotion with rein- forcement learning.Sci. Robot.9(89), eadi9579 (2024), doi: 10.1126/scirobotics.adi9579, https://www.science.org/ doi/abs/10.1126/scirobotics.adi9579

    doi:10.1126/scirobotics.adi9579 2024

  64. [65]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, I. Sutskever, Learning Transferable Visual Mod- els From Natural Language Supervision.Intl. Conf. on Ma- chine Learning (ICML)(2021)

    2021

  65. [66]

    T. L. Team, J. A. Barreiros, A. Beaulieu, A. Bhat, R. Cory, E. Cousineau, H. Dai, C. Fang, K. Hashimoto, M. Z. Ir- shad, M. Itkina, N. Kuppuswamy, K. Lee, K. Liu, D. Mc- Conachie, I. McMahon, H. Nishimura, C. Phillips-Grafflin, C. Richter, P. Shah, K. Srinivasan, B. Wulfe, C. Xu, M. Zhang, A. Alspach, M. Angeles, K. Arora, V. C. Guizilini, A. M. Castro, D...

    Pith/arXiv arXiv doi:10.48550/arxiv.2507.05331 2025

  66. [67]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Bal- akrishna, S. Nair, R. Rafailov, E. P. Foster, P. R. Sanketi, Q. Vuong, T. Kollar, B. Burchfiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, C. Finn, OpenVLA: An Open-Source Vision-Language-Action Model, inConf. on Robot Learn- ing (CoRL), P. Agrawal, O. Kroemer, W. Burgard, Eds., vol. 270 ofProceeding...

    2024

  67. [68]

    Abdolhosseini, H

    F. Abdolhosseini, H. Y. Ling, Z. Xie, X. B. Peng, M. van de Panne, On LearningSymmetricLocomotion, inSIGGRAPH, MIG ’19 (New York, NY, USA) (2019), doi:10.1145/3359566. 3360070,https://doi.org/10.1145/3359566.3360070

    arXiv 2019

  68. [69]

    Bellegarda, Y

    G. Bellegarda, Y. Chen, Z. Liu, Q. Nguyen, Robust High- Speed Running for Quadruped Robots via Deep Rein- forcement Learning, inIEEE/RSJ Intl. Conf. on Intelli- gent Robots and Systems (IROS)(2022), pp. 10364–10370, doi:10.1109/iros47612.2022.9982132, http://dx.doi.org/ 10.1109/IROS47612.2022.9982132

    arXiv 2022

  69. [71]

    R. Wang, J. Lehman, J. Clune, K. O. Stanley, Paired Open-Ended Trailblazer (POET): Endlessly Generating In- creasingly Complex and Diverse Learning Environments and Their Solutions (2019),https://arxiv.org/abs/1901. 01753

    2019

  70. [72]

    D. Chen, B. Zhou, V. Koltun, P. Krähenbühl, Learning by Cheating, inConf. on Robot Learning (CoRL), L. P. Kaelbling, D. Kragic, K. Sugiura, Eds. (PMLR), vol. 100 ofProceedings of Machine Learning Research(2019), pp. 66–75. 15 preprint for Science Robotics | ReviewReview Paper 2026

    2019

  71. [73]

    Zhuang, S

    Z. Zhuang, S. Yao, H. Zhao, Humanoid Parkour Learning, in8th Annual Conference on Robot Learning(2024),https: //openreview.net/forum?id=fs7ia3FqUM

    2024

  72. [74]

    Haarnoja, S

    T. Haarnoja, S. Ha, A. Zhou, J. Tan, G. Tucker, S. Levine, Learning to Walk Via Deep Reinforcement Learning, in Robotics: Science and Systems (RSS), A. Bicchi, H. Kress- Gazit, S. Hutchinson, Eds. (2019), doi:10.15607/RSS.2019. XV.011,https://doi.org/10.15607/RSS.2019.XV.011

    doi:10.15607/rss.2019 2019

  73. [75]

    P. Wu, A. Escontrela, D. Hafner, P. Abbeel, K. Gold- berg, DayDreamer: World Models for Physical Robot Learning, inConf. on Robot Learning (CoRL), K. Liu, D. Kulic, J. Ichnowski, Eds., vol. 205 (2022), pp. 2226–2240, https://proceedings.mlr.press/v205/wu23c.html

    2022

  74. [76]

    Haarnoja, A

    T. Haarnoja, A. Zhou, P. Abbeel, S. Levine, Soft Actor- Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor, inIntl. Conf. on Ma- chine Learning (ICML), J. G. Dy, A. Krause, Eds., vol. 80 (2018), pp. 1856–1865, http://proceedings.mlr.press/ v80/haarnoja18b.html

    2018

  75. [77]

    Abdolmaleki, J

    A. Abdolmaleki, J. T. Springenberg, Y. Tassa, R. Munos, N. Heess, M. A. Riedmiller, Maximum a Posteriori Policy Optimisation, inIntl. Conf. on Learning Representations (ICLR)(2018)

    2018

  76. [78]

    Robot.10(105), eadv3604 (2025), doi:10.1126/scirobotics.adv3604, https://www

    J.He, C.Zhang, F.Jenelten, R.Grandia, M.Bächer, M.Hut- ter, Attention-based map encoding for learning general- ized legged locomotion.Sci. Robot.10(105), eadv3604 (2025), doi:10.1126/scirobotics.adv3604, https://www. science.org/doi/abs/10.1126/scirobotics.adv3604

    doi:10.1126/scirobotics.adv3604 2025

  77. [80]

    X. B. Peng, Z. Ma, P. Abbeel, S. Levine, A. Kanazawa, AMP: adversarial motion priors for stylized physics-based character control.ACM Transactions on Graphics40(4), 1–20 (2021), doi:10.1145/3450626.3459670,http://dx.doi. org/10.1145/3450626.3459670

    arXiv 2021

  78. [81]

    Q. Liao, T. E. Truong, X. Huang, G. Tevet, K. Sreenath, C. K. Liu, BeyondMimic: From Motion Tracking to Ver- satile Humanoid Control via Guided Diffusion.CoRR abs/2508.08241(2025), doi:10.48550/ARXIV.2508.08241, https://doi.org/10.48550/arXiv.2508.08241

    Pith/arXiv arXiv doi:10.48550/arxiv.2508.08241 2025

  79. [82]

    T. He, J. Gao, W. Xiao, Y. Zhang, Z. Wang, J. Wang, Z. Luo, G. He, N. Sobanbabu, C. Pan, Z. Yi, G. Qu, K. Kitani, J. K. Hodgins, L. Fan, Y. Zhu, C. Liu, G. Shi, ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills, inRobotics: Science and Systems (RSS)(LosAngeles, CA, USA) (2025), doi: 10.15607/RSS.2025.XXI.066

    work page doi:10.15607/rss.2025.xxi.066 2025

  80. [83]

    C. Li, M. Vlastelica, S. Blaes, J. Frey, F. Grimminger, G. Martius, Learning Agile Skills via Adversarial Imita- tion of Rough Partial Demonstrations, inConf. on Robot Learning (CoRL)(PMLR), vol. 205 (2023), pp. 342–352

    2023

Showing first 80 references.