Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Optimizing Bipedal Locomotion for The 100m Dash With Comparison to Human Running

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a simulation-based search over stride frequency and swing ratio produces Cassie running gaits that closely track human running biomechanics, and that integrating these gaits into a start-to-stop controller set the…

desk verdict A credible systems result with a real hardware world record, but the headline 'similar to humans' claim is built on a deterministic simulation search whose sim-to-real gap is only partially acknowledged. read the letter →

arxiv 2508.03070 v1 pith:7EWAYYIT submitted 2025-08-05 cs.RO cs.AI

Devin Crowley , Jeremy Dao , Helei Duan , Kevin Green , Jonathan Hurst , Alan Fern This is my paper · ORCID
classification cs.ROcs.AI
keywords bipedallocomotiongaitoptimizationsim-to-realreinforcementlearningCassierobot100mdashrunningbiomechanicsgroundreactionforcecostoftransport
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

This paper asks how fast a bipedal robot like Cassie can run and whether its efficient gaits resemble human running. It answers by training one policy across a spectrum of stride-frequency and swing-ratio settings, searching in simulation for the most efficient parameters at each speed, and comparing the resulting gaits to published human biomechanics. The central claim is that the optimized speed-to-gait mapping is qualitatively different from the previous hand-tuned mapping and produces running mechanics that track human data: stride length rises and tapers, stride frequency stays flat until top speeds, and swing time is dominated by aerial phases. The paper further claims these gaits, wrapped in a start-to-stop 100m controller, set the Guinness World Record for fastest 100m by a bipedal robot at 24.37 seconds (4.10 m/s).

What carries the argument

The load-bearing object is the speed-to-gait-parameter mapping learned from a simulation search. For each speed command, gait parameters (swing ratio and stride frequency) are drawn from uniform distributions offset from the hand-tuned baseline, one policy is trained across the whole spectrum, and deterministic MuJoCo rollouts are scored by a weighted combination of speed error, cost of transport, torque, and motor velocity. The top five parameter combinations per speed define the optimized mapping. This mapping is what produces the human-like stride mechanics and what the 100m dash controller commands.

What would settle it

Run the optimized mapping on hardware and compare the actually executed swing ratio and stride frequency at each speed against the simulation-selected top five: if hardware-optimal parameters fall outside the simulation ranking or the human-like stride trends disappear, the central claim fails. A second check is to strengthen or actively control the hip-roll motors around 3 m/s; if effective ground reaction force keeps plateauing, the torque-limit explanation is wrong, and if it rises linearly like in humans, the explanation is confirmed.

Watch

Extended reading notes

Core claim

The central discovery is that an RL-trained running policy for Cassie, when its gait parameters (swing ratio and stride frequency) are optimized per speed by a simulation-based search, converges on long, infrequent steps rather than high step frequencies: the efficient gaits use a high swing ratio (lots of aerial time) with a lower stride frequency so stance time stays long enough to deliver impulse without excessive ground forces. When these optimized gaits are measured in body-agnostic units, their stride length, stride frequency, swing time, and aerial time track the human curves from [23] across 2-5 m/s. The main divergence is effective ground reaction force, which plateaus above ~3 m/s in Cassie because the hip-roll motors saturate, whereas in humans it rises roughly linearly with speed. The paper's claim is that this similarity is not incidental: it emerges from optimizing efficiency, and it supports the view that a learned speed-to-parameter mapping yields more natural, efficient running than the hand-tuned mapping.

Load-bearing premise

The whole speed-to-gait mapping is selected in a deterministic MuJoCo simulation from a single initial state; if that simulation ranks gait parameters differently from the real robot, the optimized mapping and its human-like mechanics could be artifacts of the model rather than properties of Cassie.

Editorial extensions

If this is right

  • Efficient high-speed bipedal running on Cassie favors long, infrequent steps with a pronounced aerial phase over rapid stepping.
  • Optimized gaits resemble human running mechanics in stride length, stride frequency, and swing/areal timing despite morphological differences.
  • The plateau in effective ground reaction force above about 3 m/s points to hip-roll torque saturation as a hardware limit rather than a control limit.
  • The integrated two-policy controller achieves a 24.37-second hardware time, while simulation supports speeds up to 5 m/s (about 22 seconds for 100m), a gap the paper attributes to simulator inaccuracy.
  • The hand-tuned mapping used too high a stride frequency, requiring greater ground forces; the optimized mapping avoids this by preserving stance time.

Reading between the lines

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

  • An untested extension of this result is that any bipedal runner with a similar speed range will show the same stride-length/frequency pattern when efficiency is optimized; if not, the human similarity would be coincidental to Cassie's morphology.
  • Because the simulation search used deterministic rollouts from a single initial state, a natural robustness test is to re-run the optimization over varied initial states and disturbances; the mapping may shift when averaged over conditions.
  • The sim-to-real gap near 5 m/s suggests the optimized mapping itself may not transfer intact to hardware; instrumenting the real robot to record executed swing ratio and stride frequency at each commanded speed would directly test whether the simulation-selected parameters remain optimal on hardware.
  • If hip-roll torque is indeed the limiter, a hardware redesign with stronger hip-roll actuation should push the effective ground reaction force plateau to higher speeds and shrink the gap toward simulated dash times.
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 presents a sim-to-real RL framework for optimizing two gait parameters (swing ratio and stride frequency) as a function of speed for the bipedal robot Cassie. It trains a single LSTM policy with PPO in MuJoCo over a range of speeds and gait-parameter offsets around a prior hand-tuned mapping, then selects the top-5 parameter combinations per speed using a hand-weighted cost. The paper compares the resulting simulated running mechanics to human data from Weyand et al. [23] and reports similarities in stride length, stride frequency, swing time, and aerial time, along with differences in effective ground reaction force. It finally integrates the optimized gaits into a five-stage 100m dash controller with standing-to-running and running-to-standing transitions, and reports three successful hardware trials with a best time of 24.37 s (4.10 m/s), claiming the Guinness World Record for fastest 100m by a bipedal robot.

Significance. If the claims are robust, the paper is significant: it demonstrates a complete start-to-stop 100m dash on a real bipedal robot, offers one of the first systematic speed-to-gait-parameter optimizations for a biped, and presents a transparent comparison against an independent published human dataset. The hardware record is a genuine systems achievement. However, the optimization and the human-mechanics comparison are built on deterministic simulations, and the paper itself concedes a sim-to-real gap at speeds above roughly 4 m/s, so the hardware data validate only a single operating point rather than the full speed-to-parameter mapping.

major comments (4)
  1. [Section III.B] The top-5 selection is made from deterministic rollouts that, in the paper's own words, 'used the same initial state.' No variation over initial conditions, noise, or random seeds is reported, so the cost-based ranking has no variance estimate. This is load-bearing because Figure 3 and the optimized column of Figure 4 are derived from these rankings. Please provide score distributions or stability checks over multiple initial states/random seeds, or explicitly justify why the ranking is insensitive to these choices.
  2. [Sections IV and VI] The right column of Figure 4 is computed from the simulated top-5 gaits, and Section VI states that hardware cannot reach the 5 m/s speeds the simulator supports. The abstract's claim that 'key properties of the gaits are highly similar across a wide range of speeds' is therefore unvalidated for the real robot at the upper end of the plotted range. At minimum, restrict the human-comparison claim to speeds verified on hardware, or validate the simulated efficiency ordering on hardware at multiple speeds.
  3. [Section III.B, cost function] The overall score is a hand-weighted sum of four costs, but the weights are not reported and no sensitivity analysis is given. Since three of the four terms measure efficiency, the top-5 selection could be an artifact of the chosen weighting. Please report the exact weights and show that the top-5 parameter sets are stable under reasonable weight variations.
  4. [Table I and Section VI] The hardware evidence consists of three trials, all at a 4 m/s speed command, with times spanning 24.37 to 27.38 s, and there is no hardware baseline using the hand-tuned mapping. The world-record claim is supported, but the stronger claims that the optimized mapping is 'qualitatively different' from the hand-tuned mapping and yields more natural or efficient gaits are not directly supported by hardware data. A hardware comparison of several parameter combinations across a range of speeds, or at least a sim-to-real comparison of the cost terms, is needed.
minor comments (5)
  1. [Abstract] The abstract contains a typo: 'F astest' should be 'Fastest'.
  2. [Section III.A heading] The heading 'Training Across a Spectrum of Gait Parameter V alues' has a formatting error: 'V alues' should be 'Values'.
  3. [Section III.B] The text says 'Each component is weighted to have a similar contribution' but does not give the actual formula or weights; please make the scoring metric explicit.
  4. [Section V.B] The phrase 'arbitrarily chose the median GRF crossing phase' is confusing: 'arbitrarily' suggests an unprincipled choice, whereas the surrounding text gives a rationale. Please rephrase, e.g., 'without further tuning'.
  5. [Section VI] The sentence 'none have been bridged the gap onto hardware so effectively' is ungrammatical; it should be 'none have bridged the gap onto hardware so effectively.'

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the optimized gait mapping is scored on efficiency costs that do not include human biomechanical targets, the human comparison uses external published data [23], and the hardware record is an independent demonstration.

full rationale

The paper's claimed derivation chain is not circular. The gait-parameter optimization in Section III is a simulation search over a discretized (ratio, freq) grid centered on the hand-tuned mapping of [11], and the top-5 selections are scored by four costs (speed error, cost of transport, torque, motor velocity). None of these costs includes human biomechanical targets, so the comparison in Section IV to the published human running data of [23] is an independent, post-hoc comparison rather than a fit to the target result. The hand-tuned mapping from [11] is used only as the center and offset of the search space and as a baseline; the optimized curve is free to move within that space and is reported as qualitatively different. References to the authors' prior work [4,11] supply the controller architecture, reward formulation, and baseline mapping, but these are methodological inputs rather than unverified theorems that carry the central claims. The hardware 100m dash is an external demonstration with independent witnesses. The admitted sim-to-real gap and the deterministic single-rollout search are threats to the validity or robustness of the optimization, but they are not cases where a prediction reduces by construction to an input, where a fitted parameter is renamed as a prediction, or where a uniqueness theorem is imported from self-citations. No specific circular step can be exhibited from the paper's equations or argument structure.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claims rest on the fidelity of the MuJoCo simulation and on a handful of hand-selected search and controller parameters whose values are not justified by first principles or data. No new physical entities are introduced. The comparison to human data is external and non-circular. The absence of public artifacts and the deterministic search design are the main limitations.

free parameters (4)
  • Cost weighting scheme = Equal weights across four cost components
    A hand-set weighting scheme defines the optimization objective. The paper says 'Each component is weighted to have a similar contribution to the overall score across all speeds.' Different weights would change the selected top-5 gaits and therefore the biomechanical comparison.
  • Gait parameter offset bounds = ±0.2 swing ratio, ±0.625 stride frequency
    The search region is defined by offsets from the previous hand-tuned mapping, with no principled justification given for the magnitudes. These bounds constrain the set of possible optimized gaits.
  • Median GRF crossing phase = Median of simulated distribution
    The standing-to-running transition phase is chosen by taking the median of a simulated distribution, described as 'arbitrarily chose the median GRF crossing phase' in Section V-B.
  • Median apex phase = Median per foot
    The running-to-standing transition waits for median apex phases chosen empirically. This is a hand-tuned controller detail adjusted until hardware trials succeeded.
assumptions (4)
  • domain assumption MuJoCo simulation model of Cassie with dynamics randomization is a faithful enough proxy for hardware to select optimal gaits.
    The optimization search is run entirely in simulation (Section III-B), and the paper later reports a sim-to-real speed gap at 5 m/s (Section VI). If the simulator systematically biases gait parameter rankings, the optimized mapping and the human comparison inherit that bias.
  • domain assumption The dense reward function from [11] is suitable for expressing all relevant running gaits and does not implicitly encode human-like mechanics.
    Training relies on the inherited reward; no reward weight or ablation is reported, so it is assumed to be a neutral scaffold for gait exploration.
  • domain assumption Established human biomechanics data [23] are directly comparable to Cassie in body-agnostic units and are correct for the claimed speed range.
    The comparison treats [23] as ground truth and re-plots data from that source; no independent processing or error bounds are given.
  • domain assumption Top-5 averages of deterministic simulations are representative of the gait parameter landscape.
    Search uses a single deterministic rollout per parameter combination from the same initial state; no stochastic repeats or confidence intervals are shown.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimizing Bipedal Locomotion for The 100m Dash With Comparison to Human Running." pith.science (2026). https://pith.science/paper/7EWAYYIT

@misc{pith2026250803070,
  author       = {Pith},
  title        = {Pith review of: Optimizing Bipedal Locomotion for The 100m Dash With Comparison to Human Running},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7EWAYYIT}},
  note         = {Machine review of arXiv:2508.03070}
}
read the original abstract

In this paper, we explore the space of running gaits for the bipedal robot Cassie. Our first contribution is to present an approach for optimizing gait efficiency across a spectrum of speeds with the aim of enabling extremely high-speed running on hardware. This raises the question of how the resulting gaits compare to human running mechanics, which are known to be highly efficient in comparison to quadrupeds. Our second contribution is to conduct this comparison based on established human biomechanical studies. We find that despite morphological differences between Cassie and humans, key properties of the gaits are highly similar across a wide range of speeds. Finally, our third contribution is to integrate the optimized running gaits into a full controller that satisfies the rules of the real-world task of the 100m dash, including starting and stopping from a standing position. We demonstrate this controller on hardware to establish the Guinness World Record for Fastest 100m by a Bipedal Robot.

Figures

Figures reproduced from arXiv: 2508.03070 by the authors.

Figure 1
Figure 1. The bipedal robot Cassie standing on the track after [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Hand-tuned mapping from the speed command to the [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The top 5 most efficient freq (above) and ratio (below) gait parameter commands at each commanded speed. The blue line indicates the previous hand-tuned mapping, and the curve is a 3rd degree polynomial regression. • Speed error: how well the speed command was matched • Cost of transport: the energy used / distance traveled • Torque cost: the average torque used • Motor velocity: the average motor speed Each compone… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Running mechanics as functions of speed (left) for the hand-tuned gait parameter mapping with Cassie, (center) as [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: The ground reaction force (GRF) at each policy [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 7
Figure 7. Figure 7: Speed as a function of time for a simulated 100m [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. CBF-RL: Safety Filtering Reinforcement Learning in Training with Control Barrier Functions

    cs.RO 2025-10 conditional novelty 6.0 of 10

    Training RL policies with a closed-form CBF safety filter plus CBF reward lets a Unitree G1 humanoid avoid obstacles and climb stairs without a runtime safety filter.

Reference graph

Works this paper leans on

23 extracted references · 2 linked inside Pith · cited by 1 Pith paper

  1. [23]

    Faster top running speeds are achieved with greater ground forces not more rapid leg movements,

    P. G. Weyand, D. B. Sternlight, M. J. Bellizzi, and S. Wright, “Faster top running speeds are achieved with greater ground forces not more rapid leg movements,” Journal of applied physiology , vol. 89, no. 5, pp. 1991– 1999, 2000

    1991

  2. [1]

    Reinforcement learning for robust parameterized locomotion control of bipedal robots,

    Z. Li, X. Cheng, X. B. Peng, P. Abbeel, S. Levine, G. Berseth, and K. Sreenath, “Reinforcement learning for robust parameterized locomotion control of bipedal robots,” in 2021 IEEE International Conference on Robotics and Automation (ICRA) , 2021, pp. 2811– 2817

    2021

  3. [2]

    Reinforcement learning-based cascade motion policy design for robust 3d bipedal locomotion,

    G. A. Castillo, B. Weng, W. Zhang, and A. Hereid, “Reinforcement learning-based cascade motion policy design for robust 3d bipedal locomotion,” IEEE Access, vol. 10, pp. 20 135–20 148, 2022

    2022

  4. [3]

    Learning bipedal robot locomotion from human movement,

    M. Taylor, S. Bashkirov, J. F. Rico, I. Toriyama, N. Miyada, H. Yanagisawa, and K. Ishizuka, “Learning bipedal robot locomotion from human movement,” in 2021 IEEE International Conference on Robotics and Automation (ICRA) , 2021, pp. 2797–2803

    2021

  5. [4]

    Sim- to-Real Learning of All Common Bipedal Gaits via Periodic Reward Composition,

    J. Siekmann, Y . Godse, A. Fern, and J. Hurst, “Sim- to-Real Learning of All Common Bipedal Gaits via Periodic Reward Composition,” in IEEE International Conference on Robotics and Automation (ICRA) , 2021

    2021

  6. [5]

    Blind Bipedal Stair Traversal via Sim- to-Real Reinforcement Learning,

    J. Siekmann, K. Green, J. Warila, A. Fern, and J. Hurst, “Blind Bipedal Stair Traversal via Sim- to-Real Reinforcement Learning,” in Proceedings of Robotics: Science and Systems , vol. abs/2105.08328, Virtual, 7 2021. [Online]. Available: https://arxiv.org/ abs/2105.08328

    arXiv 2021

  7. [6]

    Sim- to-real learning for bipedal locomotion under unsensed dynamic loads,

    J. Dao, K. Green, H. Duan, A. Fern, and J. Hurst, “Sim- to-real learning for bipedal locomotion under unsensed dynamic loads,” in 2022 International Conference on Robotics and Automation (ICRA) , 2022

    2022

  8. [7]

    Challenges of Learned High- Speed Locomotion over Five Kilometers in the Real World,

    J. Dao, K. Green, H. Duan, J. Siekmann, Y . Godse, A. Fern, and J. Hurst, “Challenges of Learned High- Speed Locomotion over Five Kilometers in the Real World,” in ICRA 2021: 5th Workshop on Legged Robots: Towards Real-World Deployment of Legged Robots, 2021

    2021

Show all 23 references
  1. [8]

    Fastrunner: A fast, efficient and robust bipedal robot. concept and planar simulation,

    S. Cotton, I. M. C. Olaru, M. Bellman, T. van der Ven, J. Godowski, and J. Pratt, “Fastrunner: A fast, efficient and robust bipedal robot. concept and planar simulation,” in 2012 IEEE International Conference on Robotics and Automation , 2012, pp. 2358–2364

    2012

  2. [9]

    Robust spring mass model running for a physical bipedal robot,

    W. C. Martin, A. Wu, and H. Geyer, “Robust spring mass model running for a physical bipedal robot,” in 2015 IEEE International Conference on Robotics and Automation (ICRA) , 2015, pp. 6307–6312

    2015

  3. [10]

    Bipedal robotic running with durus- 2d: Bridging the gap between theory and experiment,

    W.-L. Ma, S. Kolathaya, E. R. Ambrose, C. M. Hubicki, and A. D. Ames, “Bipedal robotic running with durus- 2d: Bridging the gap between theory and experiment,” ser. HSCC ’17. New York, NY , USA: Association for Computing Machinery, 2017, p. 265–274. [Online]. Available: https:...

    2017

  4. [11]

    Practical Reinforcement Learning for Bipedal Locomotion,

    J. Dao, “Practical Reinforcement Learning for Bipedal Locomotion,” Masters of Science in Robotics, Oregon State University, 2021

    2021

  5. [12]

    Learning loco- motion skills using deeprl: Does the choice of ac- tion space matter?

    X. B. Peng and M. van de Panne, “Learning loco- motion skills using deeprl: Does the choice of ac- tion space matter?” in Proceedings of the ACM SIG- GRAPH/Eurographics Symposium on Computer Anima- tion. ACM, 2017, p. 12

    2017

  6. [13]

    Feedback control for cassie with deep rein- forcement learning,

    Z. Xie, G. Berseth, P. Clary, J. Hurst, and M. van de Panne, “Feedback control for cassie with deep rein- forcement learning,” in 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2018, pp. 1241–1246

    2018

  7. [14]

    Sim- to-real: Learning agile locomotion for quadruped robots,

    J. Tan, T. Zhang, E. Coumans, A. Iscen, Y . Bai, D. Hafner, S. Bohez, and V . Vanhoucke, “Sim- to-real: Learning agile locomotion for quadruped robots,” in Proc. of Robotics: Science and Systems XIV. Pittsburgh, Pennsylvania: Robotics: Science and Systems Foundation, 6 2018. [...

    2018

  8. [15]

    Learning agile and dynamic motor skills for legged robots,

    J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V . Tsounis, V . Koltun, and M. Hutter, “Learning agile and dynamic motor skills for legged robots,” Science Robotics, vol. 4, no. 26, 2019. [Online]. Available: https://robotics.sciencemag.org/content/4/26/eaau5872

    2019

  9. [16]

    Deepgait: Planning and control of quadrupedal gaits using deep reinforcement learning,

    V . Tsounis, M. Alge, J. Lee, F. Farshidian, and M. Hutter, “Deepgait: Planning and control of quadrupedal gaits using deep reinforcement learning,” IEEE Robotics and Automation Letters , vol. 5, no. 2, pp. 3699–3706, 2020

    2020

  10. [17]

    Proximal Policy Optimization Algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal Policy Optimization Algorithms,” 2017. [Online]. Available: https: //arxiv.org/abs/1707.06347

    2017 arXiv

  11. [18]

    MuJoCo: A physics engine for model-based control,

    E. Todorov, T. Erez, and Y . Tassa, “MuJoCo: A physics engine for model-based control,” in 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2012, pp. 5026–5033

    2012

  12. [19]

    Learning memory-based control for human-scale bipedal locomotion,

    J. Siekmann, S. Valluri, J. Dao, L. Bermillo, H. Duan, A. Fern, and J. Hurst, “Learning memory-based control for human-scale bipedal locomotion,” in Proceedings of Robotics: Science and Systems , 7 2020

    2020

  13. [20]

    Learning quadrupedal locomotion over challenging terrain,

    J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter, “Learning quadrupedal locomotion over challenging terrain,” Science Robotics , vol. 5, no. 47, p. eabc5986, 2020. [Online]. Available: https: //www.science.org/doi/abs/10.1126/scirobotics.abc5986

    2020 doi

  14. [21]

    Sim-to-real transfer of robotic control with dynamics randomization,

    X. B. Peng, M. Andrychowicz, W. Zaremba, and P. Abbeel, “Sim-to-real transfer of robotic control with dynamics randomization,” in 2018 IEEE International Conference on Robotics and Automation (ICRA) , 2018, pp. 3803–3810

    2018

  15. [22]

    Fast and efficient locomotion via learned gait tran- sitions,

    Y . Yang, T. Zhang, E. Coumans, J. Tan, and B. Boots, “Fast and efficient locomotion via learned gait tran- sitions,” in Conference on Robot Learning . PMLR, 2022, pp. 773–783

    2022

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.