Pith. sign in

REVIEW 4 major objections 5 minor 59 references

Robust control for multi-legged elongate robots in noisy environments

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Multi-legged elongate robots have an emergent stride ceiling near four times leg length—about 33 cm per cycle—no matter how many legs are added.

desk verdict A serious, experimentally grounded robotics paper with a clean geometric speed bound; the MIMO 'approaching the bound' claim is slightly overstated due to a footnote they acknowledge but don't correct. read the letter →

arxiv 2506.15788 v1 pith:DTL5VTOF submitted 2025-06-18 cs.RO cs.SYeess.SY

Baxi Chong , Juntao He , Daniel Irvine , Tianyu Wang , Esteban Flores , Daniel Soto , Jianfeng Lin , Zhaochen Xu
show 3 more authors
Vincent R Nienhusser Grigoriy Blekherman Daniel I. Goldman
This is my paper · ORCID
classification cs.ROcs.SYeess.SY
keywords multi-leggedrobotselongateterrainrobustnessmechanicalintelligencegeometricmechanicscommunicationtheoryanalogybasicactivecontactspeedupperbound
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that multi-legged elongate robots (MERs), built from serially connected bipedal modules, have an emergent upper bound on stride length of roughly four times leg length, about 33 cm per cycle for the tested robots, no matter how many legs are added. It argues this ceiling follows from treating each leg-ground contact as a bit-like 'basic active contact' and applying communication-theory ideas: redundancy gives open-loop robustness on rough terrain, while feedback control plays the role of retransmission. Using geometric mechanics to design body-undulation gaits, the authors show that adding legs only helps up to about seven leg pairs, after which speed saturates. They then demonstrate that a learned multi-input feedback controller reaches 0.40 ± 0.03 body lengths per cycle on noisy terrain, close to the predicted 0.43, and that C-shaped legs reach comparable speeds with no feedback at all. The wider value is a design rule: speed, robustness, and sensor/computation load can be traded off systematically rather than tuned robot-by-robot.

What carries the argument

The load-bearing object is the 'bac'—basic active contact, one leg-ground contact treated as a bit in a communication channel—together with the analogy that maps mechanical intelligence (passive gravity redistribution of ground reaction forces) to a majority-vote decoder, simple repetition/spatial synchronization to repetition coding, and feedback control (computational intelligence) to automatic repeat request. Gait design is carried by geometric mechanics: the body shape space with local connection matrix A(w), whose curl (the forward height function) gives the net stride from a closed gait path via Stokes' theorem. A vertical body wave modulates contact timing to spread thrust uniformly in time, which sets the 'majority vote window size' and is the tuning knob for the speed-robustness trade-off. The claimed 4l upper bound is derived from two elementary models—non-slip kinematics and steady slipping with Coulomb friction—so the bound is a geometric consequence of leg length rather than of the friction model.

What would settle it

A robot with more legs than the saturation point that achieves a flat-ground stride strictly greater than four times its leg length (4l) would falsify the claimed bound. Separately, a rugose terrain whose height variations are strongly correlated—one obstacle lifting several legs at once—on which open-loop redundancy loses its predicted reliability would falsify the independent-bit noise model.

Watch

Extended reading notes

Core claim

For an elongate robot whose legs all do the same synchronized thrust, additional legs add reliability but not speed; the paper's central discovery is that this speed ceiling is not an engineering accident but a geometric bound. With a lateral undulation wave and one leg pair per module, geometric mechanics predicts that stride length is maximized at a body amplitude set by self-collision constraints, that the optimal amplitude falls as leg count rises, and that the resulting maximum stride converges to 4l (four times leg length, 33.6 cm for l = 8.4 cm). The saturation occurs near N = 7 leg pairs for the tested geometry. The same 4l limit is recovered from a non-slip kinematic model and from a steady-state slipping model with isotropic Coulomb friction, so it does not depend on the no-slip assumption. The authors then show the limit is approachable on rough terrain: a reinforcement-learned MIMO controller, which adapts body undulation, leg amplitude, and vertical wave in response to measured bac loss, reaches 0.40 ± 0.03 BL/cycle against the predicted 0.43, and point-foot robots with C-shaped legs reach comparable speed in open loop because the distributed foot contact makes bac loss rare.

Load-bearing premise

The load-bearing premise is that terrain-induced failures of individual leg-ground contacts behave like independent random bit errors, so that redundancy and coding arguments (majority vote, forward error correction, retransmission) apply quantitatively to locomotion.

Editorial extensions

If this is right

  • Adding legs beyond roughly seven pairs does not increase flat-ground stride; robots should be designed with fewer legs plus body undulation or feedback.
  • A simple single-input single-output controller can give a 12-legged robot reliability comparable to a 16-legged open-loop robot, showing that sensor complexity can substitute for leg count.
  • C-shaped legs let an MER traverse flat and rugose terrains with statistically indistinguishable open-loop speeds, approaching the four-times-leg-length bound.
  • The speed-robustness trade-off is controlled by encoding choices: less spatial redundancy or a smaller vertical wave raises smooth-terrain speed but costs robustness.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: terrain noise with correlated bac losses (one obstacle lifting several legs at once) should break the independent-bit redundancy guarantee; a terrain characterized by bac-loss correlation length would test this.
  • Beyond the paper: the 4l result bounds stride per cycle, not velocity; the natural next bound to seek is on cycle frequency, since absolute speed is stride times frequency.
  • Beyond the paper: the stated equivalence between computational and design complexity implies an iso-performance surface; measuring wall-clock or energy cost of a many-leg open-loop robot versus a few-leg feedback robot on identical terrain would quantify the exchange rate.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a communication-theoretic framework for multi-legged elongate robots (MERs), in which each leg-ground contact is a 'bac' analogous to a bit, passive gravity-driven force redistribution is mechanical intelligence (MI) analogous to a majority-vote decoder, and active feedback is computational intelligence (CI) analogous to ARQ. Using geometric mechanics, the authors derive an emergent upper bound on absolute speed of approximately 4l (about 33 cm/cycle for l = 8.4 cm), predict that speed saturates near N = 7 leg pairs, and validate these predictions numerically and experimentally on flat and rugose terrains. They then introduce vertical-wave contact modulation, a SISO adaptive controller, an RL-trained MIMO controller, and C-shaped legs, reporting speeds that approach the predicted bound on noisy terrain. The communication-theory analogy is used throughout to justify redundancy-based reliability and speed-robustness trade-offs.

Significance. If the results hold, the paper provides a rare predictive design principle for legged locomotion: a simple kinematic speed bound 4l that is independent of leg count, together with a quantitative account of when adding legs ceases to improve speed. The geometric-mechanics predictions match the flat-ground experiments, the N≈7 saturation is tested both in simulation and in hardware, and the C-leg experiments demonstrate terrain-agnostic open-loop performance. These are concrete strengths. The communication analogy is stimulating but not yet fully load-bearing: the reliability guarantees rest on an untested independence assumption for bac errors, several key proofs are deferred to the SI, and the MIMO comparison to the upper bound uses an inconsistent reference leg length. The paper also does not state data/code availability, which limits reproducibility of the simulation and RL components. With targeted revisions, this could be a strong contribution to the field.

major comments (4)
  1. [Approach the upper bound of speed on noisy landscapes; footnote 4] The claim that MIMO CI at 0.40±0.03 BL/cycle 'approaches the theoretical upper bound of 0.43 BL/cycle' is internally inconsistent with footnote 4, which states that the contact sensors effectively increase leg length and thereby raise the upper bound. Since the predicted bound is 4l, the comparison must use the effective leg length of the sensor-equipped robot. As a concrete illustration, increasing l from 8.4 cm to 9.4 cm changes the bound to approximately 0.48 BL/cycle, making the achieved speed about 83% of the bound rather than 93%. Please recompute the comparison with the actual effective leg length or rephrase the claim so that it does not overstate proximity to the bound.
  2. [Temporal and spatial synchronization for simple repetition] The main text twice asserts that a constant thrust profile minimizes the coefficient of variation of cycle-averaged velocity and that a more uniform thrust distribution lowers Cv, with proofs deferred to the SI. The SI is not supplied with the manuscript as provided, so these load-bearing claims are not verifiable from the submission. Please include the proofs or state the exact modeling assumptions under which they hold, and confirm that the vertical-wave robustness results do not depend on an unstated circularity between the thrust-uniformity measure and the performance metric.
  3. [Gravity as an MI control scheme; Table 1] The reliability framework assumes that terrain-induced bac losses behave like independent bit errors in a communication channel. The experiments use random-height step fields, but the paper does not measure or model correlated bac losses, such as a single obstacle lifting several legs at once or a body tilt causing systematic contact loss. Under correlated noise, the majority-vote and redundancy guarantees need not hold. Please either add experiments or analysis with controlled correlation lengths of terrain failures, or explicitly limit the reliability claims to the tested class of terrains.
  4. [Embedding MI in bac through foot morphology design] The C-leg result is compared to '4×leg length, 0.44 BL/cycle,' but the effective leg length of the C-shaped leg is not defined. If the C-leg's effective length differs from the point-leg value l = 8.4 cm used for the earlier bound, the reference bound changes and the 'approaching the upper bound' claim is weakened. Please specify the leg length used for the C-leg comparison and recompute the bound accordingly, or report the ratio of achieved speed to the appropriate 4l value for the C-leg geometry.
minor comments (5)
  1. [Emergent upper bound of absolute speed with additional legs] The text contains the typo 'self-collusion' where 'self-collision' is meant; please correct it.
  2. [Multi-input-multi-output CI via reinforcement learning] The phrase 'MIMI CI' appears in the text and should be 'MIMO CI'.
  3. [Table 1] The table entry 'Automatic-Repeat Qequest' contains a typo; it should read 'Automatic Repeat Request'.
  4. [Figure 7.A.2] The figure labels the non-MIMO controller 'SIMO CI,' while the text refers to it as 'SISO CI'; please unify the terminology.
  5. [Single-input-single-output CI] The SISO controller gain p = 3.2 is chosen from empirical experiments; a sensitivity analysis or cross-validation would help establish that the reported improvement is not specific to this particular value.

Circularity Check

1 steps flagged · score 5.0 of 10

The 4l speed bound is derived independently, but the N=7 saturation is an experimentally noticed crossover relabeled as a theory prediction, and the MIMO 'approaching bound' comparison uses a stale, lower bound after the paper admits the robot's leg length increased.

  1. fitted input called prediction [Section 'Emergent upper bound of absolute speed with additional legs', Eq. (6) and Fig. 4.B.1]
    "In our robot experiments, we notice that A_SC < A*_b when N<7; and A_SC > A*_b when N>7. Based on the observation, we hypothesized that increasing the number of legs can lead to substantially longer stride when N<7; however, above N=7, further increasing the number of legs will not lead to longer stride. ... the saturation N in both numerical and experimental analysis is reasonably close to 7, as predicted in theory."

    The predicted threshold value N=7 is read off the empirical crossover of A_SC and A*_b that was noticed in the robot experiments, and Eq. (6) then uses that same crossover to select the body undulation amplitude. Calling the resulting saturation 'predicted in theory' relabels an empirical observation as a first-principles prediction, and the experimental verification on the same MER platforms is not an independent test. The 4l upper bound itself is not involved in this circularity.

full rationale

The paper's headline upper bound on absolute speed (4l = 33.6 cm/cycle) is derived from a parameter-free kinematic model with non-slip and slipping Coulomb-friction arguments, not from fitting speed data, so that core derivation is self-contained and non-circular. The MI/FEC and CI/ARQ analogy is taken from the authors' prior published work, but it is used as an organizing framework rather than as an unverified uniqueness theorem, and the geometric-mechanics height-function calculations are independent computations. The N=7 saturation claim is the main partial-circularity point: the threshold was first noticed in the same robot experiments later cited as verification, although the numerical simulation provides some independent support. Separately, footnote 4 near the MIMO CI claim explicitly admits that the sensor-equipped robot has a longer effective leg length, which raises the applicable 4l_eff bound above the quoted 0.43 BL/cycle; comparing the measured 0.40 BL/cycle against the stale 0.43 bound is therefore not established by the paper's own equations. That is a load-bearing comparison error rather than a definitional circularity, and it weakens the 'approaching the upper bound' narrative without invalidating the geometric bound itself.

Assumptions & free parameters 4 free parameters · 5 assumptions · 1 invented entities

The central claims rest on five modeling assumptions (quasi-statics, Coulomb friction, independent bac noise, kinematic bound assumptions, and a restricted shape space) and four fitted or empirical parameters (p, desired duty factor, RL policy weights, self-collision threshold). The quasi-static and friction assumptions are validated on flat ground; the independent-noise assumption is the most fragile because it underpins the entire communication-theoretic reliability argument.

free parameters (4)
  • SISO controller gain p = 3.2 rad
    Chosen from empirical experiments to set adaptive vertical wave amplitude Ap = p(d - d_hat). Affects the claimed improvement of adaptive over open-loop control.
  • Desired duty factor d = not reported
    Reference setpoint in the SISO controller; no numerical value is given in the text.
  • MLP policy weights (MIMO CI) = not reported (trained via RL)
    The multilayer perceptron is trained by reinforcement learning, with details deferred to ref (47). These weights encode the MIMO control policy that yields the claimed 0.40 BL/cycle speed.
  • Self-collision threshold = 1 cm foot separation
    Configurations with adjacent feet closer than 1 cm are counted as self-collision; this threshold sets ASC in Eq. (6), which directly determines the predicted saturation leg number N≈7.
assumptions (5)
  • domain assumption The robot operates in a quasi-static, thrust-dominated regime (coasting number C ≈ 0.01), so inertia is negligible and force balance (Eq. 2) holds at every instant.
    Invoked before Eq. (2) in 'Encoding with geometric mechanics'. If inertia were significant, the connection-vector-field model (Eq. 3-5) and all derived speed predictions would fail.
  • domain assumption Foot-ground contact follows isotropic Coulomb dry friction, with ground reaction force of fixed magnitude opposite to the slip direction.
    Used in the GRF analysis behind Eq. (2) and Fig. 3.B. Real foam-block terrain may deviate from this idealization, though flat-ground predictions match experiments.
  • domain assumption Terrain-induced bac losses behave like independent bit errors in a communication channel, so redundancy (FEC/MI) and retransmission (ARQ/CI) analogies apply.
    This is the foundational analogy in the abstract and Table 1. If terrain failures are correlated (e.g., a block lifts several legs at once), majority-vote decoding and the reliability guarantees developed in the paper lose their quantitative foundation.
  • domain assumption For the 4l speed bound, the non-slip model assumes no foot slip and one contralateral leg always in stance; the slipping model assumes steady-state equal forward/backward slip (d1 = d2).
    Used in 'Source entropy of multi-legged locomotion' to derive the kinematic upper bound. The steady-state slip assumption is plausible but not rigorously proven for a driven dissipative system.
  • ad hoc to paper The useful gait space is parameterized by two Fourier modes (Eq. 1) and circular gait paths in shape space, with all other shape coordinates prescribed by prior lower-level controllers.
    The GM analysis and optimal gait searches are restricted to this 2D subspace. The predictions (N≈7 saturation, 4l approach) are only shown for these gaits; richer shape variations might behave differently.
invented entities (1)
  • Bac (basic active contact)
    purpose: A single leg-ground contact treated as a unit of thrust generation, analogous to a bit in signal transmission.
    This is a descriptive abstraction of existing contact mechanics, not a new physical object. Its empirical handle is leg-contact observation, but it does not constitute a falsifiable entity outside the paper's own framework.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust control for multi-legged elongate robots in noisy environments." pith.science (2026). https://pith.science/paper/DTL5VTOF

@misc{pith2026250615788,
  author       = {Pith},
  title        = {Pith review of: Robust control for multi-legged elongate robots in noisy environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DTL5VTOF}},
  note         = {Machine review of arXiv:2506.15788}
}
read the original abstract

Modern two and four legged robots exhibit impressive mobility on complex terrain, largely attributed to advancement in learning algorithms. However, these systems often rely on high-bandwidth sensing and onboard computation to perceive/respond to terrain uncertainties. Further, current locomotion strategies typically require extensive robot-specific training, limiting their generalizability across platforms. Building on our prior research connecting robot-environment interaction and communication theory, we develop a new paradigm to construct robust and simply controlled multi-legged elongate robots (MERs) capable of operating effectively in cluttered, unstructured environments. In this framework, each leg-ground contact is thought of as a basic active contact (bac), akin to bits in signal transmission. Reliable locomotion can be achieved in open-loop on "noisy" landscapes via sufficient redundancy in bacs. In such situations, robustness is achieved through passive mechanical responses. We term such processes as those displaying mechanical intelligence (MI) and analogize these processes to forward error correction (FEC) in signal transmission. To augment MI, we develop feedback control schemes, which we refer to as computational intelligence (CI) and such processes analogize automatic repeat request (ARQ) in signal transmission. Integration of these analogies between locomotion and communication theory allow analysis, design, and prediction of embodied intelligence control schemes (integrating MI and CI) in MERs, showing effective and reliable performance (approximately half body lengths per cycle) on complex landscapes with terrain "noise" over twice the robot's height. Our work provides a foundation for systematic development of MER control, paving the way for terrain-agnostic, agile, and resilient robotic systems capable of operating in extreme environments.

Figures

Figures reproduced from arXiv: 2506.15788 by the authors.

Figure 3
Figure 3. C.2. Emergent upper bound of absolute speed with additional legs We next use geometric mechanics to develop an encoding scheme for MERs with different numbers of legs. Following similar processes, we numerically obtained the forward height functions and tested the modeling predictions on the robot experiments. We illustrated forward height functions and experimental results with N = 7 and N = 3 in [PITH_FULL_IMAGE:… view at source ↗
Figure 4
Figure 4. B.1). That being said, with better encoding scheme, we speculate that it is possible to [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 7
Figure 7. A.1) in control parameters (Fig.7.A). For details on the machine learning model and [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figures from the paper (7 more)
Figure 1
Figure 1. Figure 1: Active and passive coding for MERs (A) (top) A side-view picture of a 10-legged MER locomoting on a rugose terrain. White arrow points to the leg with a bac (basic active contact) loss. (bottom) An illustration of (left) a bac and (right) a terrain-induced bac loss. (B…
Figure 2
Figure 2. Figure 2: Trade-off between robustness and speed (A.1) Column shows the trade-off between speed and robustness as a function of design redundancy for a fixed control. Row shows the trade-off between speed and robustness as a function of control redundancy for a given robot. (A.2…
Figure 3
Figure 3. Figure 3: Geometry of MER locomotion (A.1) The shape space for body undulation of a 12-legged MER. Each point in the shape space prescribes a specific body posture (including body/leg joint angles and contact distribution). (A.2) (top) One body posture in the shape space, corres…
Figure 4
Figure 4. Figure 4: Upper bound on absolute velocity on flat ground (A) Body undulation analysis for (A.1) a 14-legged MER and (A.2) a 6-legged MER respectively. (left) Shape space with numerically computed self-collision region labeled in blue shade. (mid) The forward height function. (r…
Figure 5
Figure 5. Figure 5: Contact modulation to counter terrain rugosity (A.1) An illustrations of thrust generation from bacs (basic active contact). (A.2) The instantaneous thrust (unit: µN, µ is fric￾tion coefficient and N is the normal supporting force) as a function of time, derived from (…
Figure 6
Figure 6. Figure 6: Active coding to improve locomotion performance on terrains with varying ru￾gosity. (A) The use of a contact sensor to detect bac loss. (B) Rendering of a rugose terrain with “simple” and “complex” parts of the terrain labeled. (C) The simple linear single-input-single…
Figure 7
Figure 7. Figure 7: Leaning multi-input-multi-output (MIMO) CI for MER locomotion. (A.1) The MIMO CI model. The input variables are current body lateral and vertical amplitude (Ab(t) and Ap(t)), current leg amplitude (Aleg(t)), and the measured average duty factor ( ˆd). The output variab…

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

59 extracted references · 57 canonical work pages

  1. [1]

    R. M. Alexander,Principles of animal locomotion(Princeton University Press, 2003)

    work page 2003

  2. [2]

    Childress, A

    S. Childress, A. Hosoi, W. W. Schultz, J. Wang,Natural locomotion in fluids and on sur- faces: swimming, flying, and sliding, vol. 155 (Springer, 2012)

    work page 2012

  3. [3]

    Gravish, G

    N. Gravish, G. V . Lauder,Journal of Experimental Biology221, jeb138438 (2018)

    work page 2018

  4. [4]

    S. Kim, C. Laschi, B. Trimmer,Trends in biotechnology31, 287 (2013)

    work page 2013

  5. [5]

    Aguilar,et al.,Reports on Progress in Physics79, 1 (2016)

    J. Aguilar,et al.,Reports on Progress in Physics79, 1 (2016)

    work page 2016

  6. [6]

    Weber, E

    W. Weber, E. Weber,Verlag der Dietrichschen Buchhandlung, G¨ottingen(1836)

    work page

  7. [7]

    J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, M. Hutter,Science robotics5, eabc5986 (2020)

    work page 2020

  8. [8]

    M. H. Raibert,Legged robots that balance(MIT press, 1986)

    work page 1986

Show all 59 references
  1. [9]

    Tan,et al.,arXiv preprint arXiv:1804.10332(2018)

    J. Tan,et al.,arXiv preprint arXiv:1804.10332(2018)

    2018 arXiv

  2. [10]

    Raibert, K

    M. Raibert, K. Blankespoor, G. Nelson, R. Playter,IFAC Proceedings Volumes41, 10822 (2008)

    2008

  3. [11]

    Hutter,et al.,2016 IEEE/RSJ international conference on intelligent robots and systems (IROS)(IEEE, 2016), pp

    M. Hutter,et al.,2016 IEEE/RSJ international conference on intelligent robots and systems (IROS)(IEEE, 2016), pp. 38–44

    2016

  4. [12]

    Rudin, D

    N. Rudin, D. Hoeller, P. Reist, M. Hutter,Conference on Robot Learning(PMLR, 2022), pp. 91–100

    2022

  5. [13]

    Huang, J

    J.-K. Huang, J. W. Grizzle,IEEE Transactions on Robotics39, 2093 (2023). 34

    2023

  6. [14]

    Sitti,Extreme Mechanics Letters46, 101340 (2021)

    M. Sitti,Extreme Mechanics Letters46, 101340 (2021)

    2021

  7. [15]

    Wang,et al.,Science Robotics8, eadi2243 (2023)

    T. Wang,et al.,Science Robotics8, eadi2243 (2023)

    2023

  8. [16]

    Saranli, M

    U. Saranli, M. Buehler, D. E. Koditschek,The International Journal of Robotics Research 20, 616 (2001)

    2001

  9. [17]

    Ozkan-Aydin, B

    Y . Ozkan-Aydin, B. Chong, E. Aydin, D. I. Goldman,2020 3rd IEEE International Confer- ence on Soft Robotics (RoboSoft)(IEEE, 2020), pp. 156–163

    2020

  10. [18]

    Chong,et al.,Bioinspiration & Biomimetics17, 046015 (2022)

    B. Chong,et al.,Bioinspiration & Biomimetics17, 046015 (2022)

    2022

  11. [19]

    Chong,et al.,Proceedings of the National Academy of Sciences120, e2213698120 (2023)

    B. Chong,et al.,Proceedings of the National Academy of Sciences120, e2213698120 (2023)

    2023

  12. [20]

    Chong,et al.,Science380, 509 (2023)

    B. Chong,et al.,Science380, 509 (2023)

    2023

  13. [21]

    J. R. Pierce,An introduction to information theory: symbols, signals and noise(Courier Corporation, 2012)

    2012

  14. [22]

    Chong,et al.,Bulletin of the American Physical Society(2023)

    B. Chong,et al.,Bulletin of the American Physical Society(2023)

    2023

  15. [24]

    J. Y . Jun, D. Haldane, J. E. Clark,Experimental Robotics: The 12th International Sympo- sium on Experimental Robotics(Springer, 2014), pp. 759–773

    2014

  16. [25]

    Wilczek, A

    F. Wilczek, A. Shapere,Geometric phases in physics, vol. 5 (World Scientific, 1989)

    1989

  17. [26]

    Shapere, F

    A. Shapere, F. Wilczek,Am. J. Phys57, 514 (1989)

    1989

  18. [27]

    Shapere, F

    A. Shapere, F. Wilczek,Physical Review Letters58, 2051 (1987)

    1987

  19. [28]

    Berry,Physics Today43, 34 (1990)

    M. Berry,Physics Today43, 34 (1990). 35

    1990

  20. [29]

    Montgomery,Comm

    R. Montgomery,Comm. Math. Phys.128, 565 (1990)

    1990

  21. [30]

    J. E. Marsden, R. Montgomery, T. S. Rat,iu,Reduction, symmetry, and phases in mechanics, vol. 436 (American Mathematical Soc., 1990)

    1990

  22. [31]

    P. S. Krishnaprasad, D. P. Tsakiris,Proceedings of 1994 33rd IEEE Conference on Decision and Control(1994), vol. 3, pp. 2955–2960 vol.3

    1994

  23. [32]

    S. D. Kelly, R. M. Murray,Journal of Robotic Systems12, 417 (1995)

    1995

  24. [33]

    Ostrowski, J

    J. Ostrowski, J. Burdick,The international journal of robotics research17, 683 (1998)

    1998

  25. [34]

    A. D. Lewis,IEEE Transactions on Automatic Control45, 1420 (2000)

    2000

  26. [35]

    A. M. Bloch,Nonholonomic Mechanics(Springer New York, New York, NY , 2003), chap. 5, pp. 207–276

    2003

  27. [36]

    J. B. Melli, C. W. Rowley, D. S. Rufat,SIAM Journal on Applied Dynamical Systems5, 650 (2006)

    2006

  28. [37]

    C. E. Shannon,The Bell system technical journal27, 379 (1948)

    1948

  29. [38]

    E. F. Moore, C. E. Shannon,Journal of the Franklin Institute262, 191 (1956)

    1956

  30. [39]

    J. M. Rieser,et al.,Proceedings of the National Academy of Sciences121, e2320517121 (2024)

    2024

  31. [40]

    R. J. Full, D. E. Koditschek,Journal of Experimental Biology202, 3325 (1999)

    1999

  32. [41]

    J. M. Rieser,et al.,arXiv preprint arXiv:1906.11374(2019)

    2019 arXiv

  33. [42]

    Chong, T

    B. Chong, T. Wang, E. Erickson, P. J. Bergmann, D. I. Goldman,Proceedings of the Na- tional Academy of Sciences119, e2118456119 (2022). 36

    2022

  34. [43]

    R. M. Murray, Z. Li, S. S. Sastry, S. S. Sastry,A mathematical introduction to robotic manipulation(CRC press, 1994)

    1994

  35. [44]

    R. L. Hatton, H. Choset,The European Physical Journal Special Topics224, 3141 (2015)

    2015

  36. [45]

    D. H. Wolpert, W. G. Macready,IEEE transactions on evolutionary computation1, 67 (1997)

    1997

  37. [46]

    A. S. Tanenbaum,ACM Computing Surveys (CSUR)13, 453 (1981)

    1981

  38. [47]

    J. He, B. Chong, Z. Xu, S. Ha, D. I. Goldman,arXiv preprint arXiv:2409.09473(2024)

    2024 arXiv

  39. [48]

    J. C. Spagna, D. I. Goldman, P.-C. Lin, D. E. Koditschek, R. J. Full,Bioinspiration & biomimetics2, 9 (2007)

    2007

  40. [49]

    Altendorfer,et al.,Autonomous Robots11, 207 (2001)

    R. Altendorfer,et al.,Autonomous Robots11, 207 (2001)

    2001

  41. [50]

    He,et al.,IEEE Transactions on Robotics (submitted)(2024)

    J. He,et al.,IEEE Transactions on Robotics (submitted)(2024)

    2024

  42. [51]

    Flores, B

    E. Flores, B. Chong, D. Soto, D. I. Goldman,arXiv preprint arXiv:2410.01050(2024)

    2024 arXiv

  43. [52]

    Iaschi,et al.,arXiv preprint arXiv:2410.01046(2024)

    M. Iaschi,et al.,arXiv preprint arXiv:2410.01046(2024)

    2024 arXiv

  44. [53]

    Accessed: 2025-05-29

    Icra 2025 quadruped robot challenges,https:// quadruped-robot-challenges.notion.site/ ICRA-2025-Quadruped-Robot-Challenges-698cbe44ec434512a0064d3799f18c85. Accessed: 2025-05-29

    2025

  45. [54]

    D. E. Koditschek, R. J. Full, M. Buehler,Arthropod structure & development33, 251 (2004)

    2004

  46. [55]

    Butler, K

    Z. Butler, K. Kotay, D. Rus, K. Tomita,Proceedings 2002 IEEE International Conference on Robotics and Automation (Cat. No. 02CH37292)(IEEE, 2002), vol. 1, pp. 809–816. 37

    2002

  47. [56]

    Yasui,et al.,Scientific reports9, 1 (2019)

    K. Yasui,et al.,Scientific reports9, 1 (2019)

    2019

  48. [57]

    Gilbert,Journal of Comparative Physiology A181, 217 (1997)

    C. Gilbert,Journal of Comparative Physiology A181, 217 (1997)

    1997

  49. [58]

    A. J. Ijspeert,Annual Review Of Control, Robotics, And Autonomous Systems, Vol 3, 2020 3, 173 (2020)

    2020

  50. [59]

    Sponberg,Physics Today70, 34 (2017)

    S. Sponberg,Physics Today70, 34 (2017)

    2017

  51. [60]

    Baines,et al.,Nature610, 283 (2022)

    R. Baines,et al.,Nature610, 283 (2022). 38

    2022

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.