Pith. sign in

REVIEW 3 major objections 7 minor 94 references

Bicycle Acrobatics with Reinforcement Learning

T0 review · 3 major / 7 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 The paper claims that reinforcement learning lets a bicycle robot learn and chain jumps, flips, kips, and wheelies, validated on hardware in repeated long repertoires.

desk verdict Genuine hardware advance in bicycle acrobatics with RL, but the robustness evidence needs tighter reporting before publication. read the letter →

arxiv 2608.00880 v1 pith:GGTAT3WT submitted 2026-08-01 cs.RO

Shamel Fahmi , Arianna Ilvonen , Samuel Zapolsky , Yu-Ming Chen , Ravi Boggavarapu , Paul Drews , Ashwin Khadke , Dean Molinaro
show 3 more authors
Kaiyu Zheng Alfred Rizzi Gabriel Nelson
This is my paper · ORCID
classification cs.RO
keywords reinforcementlearningbicyclerobotacrobaticstuntsmotionimitationwaypointfollowingpolicyorchestrationsim-to-realtransfernon-holonomicdynamics
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

Reinforcement learning turns a bicycle robot into an acrobat. This paper claims that a bicycle-style robot—unstable by design and harder to control than a legged machine—can learn jumps, flips, kips, wheelies, and lateral hops in simulation and then execute them on real hardware, repeatedly and in long chained sequences. The key move is modular: each stunt is trained as its own policy, and an orchestrator switches between policies only when the robot's state satisfies the next stunt's entry conditions. If the claim holds, wheeled locomotion gains much of the agility previously reserved for legged robots, while keeping the speed and efficiency of wheels.

What carries the argument

Two mechanisms carry the argument. First, a set of specialist RL policies, each trained with one of five formulations ranging from sparse waypoint following to dense motion imitation; the waypoint-following policy sees a robot-centric heightmap extracted from a global terrain map, so jumping emerges from reaching goals rather than being scripted. Second, an orchestrator: a finite state machine built on sequential composition, in which each stunt is a node and a transition fires only when the robot's current state lies in the next policy's region of attraction. The orchestrator is what turns isolated stunts into long repertoires, because it waits for safe, state-dependent conditions instead of switching on a timer or button alone.

What would settle it

Run the waypoint-following jump experiment with the motion-capture feed disabled and the offline heightmap withheld, so the robot must localize itself and perceive the table from onboard sensors only; if the robot still completes the single- and multi-table jumps, the paper's environmental-support premise is wrong, and if it cannot, the autonomy claim is bounded to instrumented environments.

Watch

Extended reading notes

Core claim

The paper's central claim is that reinforcement learning can give a bicycle robot a repertoire of dynamic acrobatic stunts that transfer from simulation to a physical platform, and that these stunts can be chained into extended, repeatable performances. On the Ultra Mobility Vehicle (UMV), a 23.5 kg single-track robot, individual policies trained with waypoint following, pose reaching, twist tracking, guided tracking, and motion imitation produce table jumps up to 1 m, front flips, kip-ups and kip-downs, wheelies, bunny hops, three-point turns, and lateral jumps. A state-driven orchestrator sequences these policies, and the paper reports more than 15 consecutive autonomous waypoint-following jumps, more than 20 consecutive kip-jump-flip-kip-down repertoires, and more than 10 consecutive wheelie-lateral-jump repertoires on hardware. The authors read these results as evidence that RL can give wheeled robots a level of agility previously associated mainly with legged platforms.

Load-bearing premise

The autonomy results are anchored to the laboratory: the robot localizes through an external motion-capture system and reads terrain from an offline map generated before each experiment, because it has no onboard camera or other exteroceptive sensor; if that instrumentation is removed, the waypoint-following jumps and multi-table adaptations would not transfer.

Editorial extensions

If this is right

  • Wheeled robots can cross obstacles that previously required legged platforms, since the same policy clears 75 cm and 1 m tables and composes jumps with flips and kips.
  • One jump policy generalizes beyond its training distribution: it handles two-table configurations it never saw and adapts to table heights up to 1 m without retraining.
  • New stunts can be added to a repertoire by training a specialist policy and defining its entry and exit conditions, without retraining existing behaviors.
  • RL discovers energy- and hardware-friendly strategies, such as the snake-like climb before takeoff and the brief wheelie before descent, that a designer would be unlikely to specify by hand.
  • Repeated long trials (20 or more kip-jump-flip cycles) suggest the policies recover from landing impacts and maintain robustness over continuous operation, not just in single demonstrations.

Reading between the lines

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

  • Removing the motion-capture and offline-map instrumentation would likely collapse the waypoint-following autonomy; the paper's results therefore do not yet imply outdoor or unstructured-environment operation until onboard perception is added.
  • The orchestrator's hand-designed state boundaries and fixed transition timings (for example, 0.8 seconds after a trigger) may limit scalability; learning transition conditions from data is a natural extension.
  • Because the five RL formulations are morphology-agnostic, the same pipeline could be retrained for scooters, motorcycles, or cargo bikes, with the main effort going into new reference trajectories and reward shaping.
  • The low success rates on multi-table jumps hint that purely reactive, memoryless policies are near the limit of what this architecture can do; the paper's proposed high-level motion generator plus low-level tracker is a testable remedy.
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

3 major / 7 minor

Summary. This paper presents a reinforcement-learning framework for a custom bicycle robot (UMV) and demonstrates a diverse set of acrobatic stunts, including table jumps, flips, kips, wheelies, lateral jumps, bunny hops, and three-point turns. The authors train specialist policies with five different RL formulations (waypoint following, pose reaching, SE2 twist tracking, guided tracking, and motion imitation) and introduce an FSM-based orchestrator that transitions between policies using state-dependent triggers to execute long-horizon repertoires. They report hardware validation with repeated runs, including more than 15 consecutive autonomous jumps, more than 20 consecutive kip-jump-flip-kip-down repertoires, and more than 10 consecutive wheelie-lateral-jump repertoires, alongside simulation ablations in the supplementary material. The central claim is that RL can give bicycle robots a level of agility previously associated with legged platforms.

Significance. If the reported results hold, this is a notable empirical advance: it demonstrates a large repertoire of dynamic, long-horizon acrobatic behaviors on a real underactuated, non-holonomic platform, with repeated hardware trials, MoCap traces, ablations, and careful disclosure of the state-estimation infrastructure (S8). The paper also provides useful evidence on curriculum design, reward engineering, and orchestrator-based policy composition. However, the quantitative evidence for the central 'robust execution' claim is currently unauditable: Table 1 success rates lack denominators and trial definitions, and the Discussion's 'without failures' statement is inconsistent with several non-unit rates. The S5 bailout disclosure further clouds the 'autonomous consecutive repertoires' headline. The absence of released code, data, or policy weights is an additional reproducibility gap.

major comments (3)
  1. [Table 1 and Discussion] The paper's robustness claim rests on success rates that are not auditable. Table 1 reports success rates such as 0.67 (row 3), 0.50 (row 5), and 0.97 (row 21) with no trial counts, denominators, confidence intervals, or a definition of what counts as a trial. The Discussion states 'We ran these stunts several times on several robots without failures', which is inconsistent with the non-unit success rates unless the rates were computed on a different condition (e.g., simulation or non-consecutive attempts). Please report the number of attempts and successes for every Table 1 entry, state explicitly whether each rate is hardware or simulation, and reconcile the 'without failures' claim with the reported rates.
  2. [Section S5 and Table 1 rows 22–25] The emergency 'bail out' disclosure directly affects the counted successes behind the abstract's 'more than 10 consecutive autonomous and steerable repertoires'. Section S5 states that if the robot is noticed to be about to fail or fall, the orchestrator transitions to a stabilizing policy before proceeding with the stunt. If any of the consecutive trials counted toward the headline included such a bailout, the claim overstates what the policies alone achieved. Please state whether bailouts occurred in the counted trials, and if so, report the number of trials completed without any human-triggered intervention separately from those with bailouts.
  3. [Data, code, and materials availability] The availability statement says 'Data and figure generation code for this study will be available if approved by our internal review process', meaning hardware logs, policy weights, and code are not released. Since the central claims are empirical and hardware-specific, the absence of released trial logs makes it impossible for a reader to verify the success counts or the sim-to-real details. Please either release the data/code or provide a detailed raw-data appendix of all hardware trials (including failures) in the supplementary material.
minor comments (7)
  1. [Figure 4(B)] The radar chart is described as normalized, but the normalization method is not defined; please specify how each metric is scaled.
  2. [Results section] The text contains missing references/placeholders, e.g., 'All the aforementioned experiments ... are shown in .' and 'show multiple trials.' These need to be filled or the text removed.
  3. [Section S8] The statement that the robot 'also supports a fully onboard estimator for outdoor operation [1]' should be reconciled with the statement that 'the current platform lacks onboard exteroceptive sensors'; clarify whether the outdoor estimator assumes additional sensors not present in the current platform.
  4. [Section S4 and Figure S2] The ablation success rates are simulation results; please label them as simulation in the main text and figure captions to avoid confusion with the hardware success rates in Table 1.
  5. [Table S1(A)] The noise magnitudes for 'Bike Global Position' (N(0,0.1)) are given without units; add units (e.g., meters).
  6. [Orchestrator section] The phrase 'We designed the orchestrator as an Finite State Machine' should read 'a Finite State Machine'.
  7. [References [14] and [15]] References [14] and [15] are identical; please correct or replace the duplicate.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the hardware stunts are independent external benchmarks, and no reported success measure is reconstructed from a fitted parameter or from a self-citation.

full rationale

The paper's derivation chain is empirical rather than definitional. Each RL policy is trained in simulation from stated rewards, constraints, curricula, and domain randomization, then deployed on the UMV, and the claimed stunts are judged by whether the physical robot clears the table, completes the flip, or finishes the repertoire. These outcomes are external events; they are not quantities that the training objectives or reward weights reproduce by construction. The citations to same-group prior work (UMV system design [1], IMI [65], LineRides [68]) supply hardware and training methodology, but the hardware demonstrations are separately shown in the figures and movies, so the central agility claim does not reduce to those citations. The orchestrator is an FSM based on the external sequential-composition framework [74], with state-dependent transition conditions that are not fitted to the reported success counts. The S5 disclosure of a human-triggered emergency 'bail out' policy for the wheelie/lateral-jump repertoires is a genuine evidence-quality concern: it may mean that some counted 'autonomous' successes included operator intervention, and the Discussion's 'without failures' statement is hard to reconcile with the 0.50-0.97 rates in Table 1. However, these are ambiguity and auditability problems about what was measured, not circularity: no equation in the paper is shown to equal its own input, and no fitted parameter is renamed as a prediction. The paper is therefore self-contained against external benchmarks, with no demonstrated circular step.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities, conserved quantities, or theoretical constructs. The Yoll is a mechanical joint on the physical robot, and the orchestrator is a software composition based on published sequential-composition ideas. The load-bearing inputs are numerous hand-tuned parameters and domain assumptions about simulation fidelity, external localization, funnel overlaps, and reward shaping.

free parameters (8)
  • Reward weights for each training type = Various; e.g., distance-to-goal weight 6, imitation orientation weight 20 (Tables S3-S7)
    Hand-tuned and curriculum-adapted weights shape each learned behavior and directly affect which stunts emerge and their success rates.
  • Action scales, stiffness, and damping per task = e.g., head action scale 0.1 nominal vs 1.5 in pose reaching; head stiffness 120-200; damping 5-10 (Table S2)
    Chosen per task based on empirical tuning; these PD setpoint parameters determine the effective policy dynamics and sim-to-real behavior.
  • PPO entropy coefficients per task = 0.006, 0.0035, 0.0035, 0.002, 0.0025 (Table S12)
    Set manually by task complexity to balance exploration vs exploitation; affects convergence and final policy quality.
  • Curriculum iteration thresholds = e.g., 5000 iterations for waypoint-following termination curricula; varied per task (Table S10)
    Hand-chosen schedules for when reward and termination terms activate; influence learning stability and final behavior.
  • Wheel contact stiffness and damping = Not reported numerically; 'sampled many values and picked the ones that matched' drop-test data (Section S6)
    Selected to match physical drop-test force and position profiles; a manual fit to data that is load-bearing for sim-to-real transfer.
  • PD gains and torque constants = Torque constants experimentally identified; PD gains tuned by sinusoidal identification (Table S2, Section S7)
    System identification on the physical robot; the specific values are empirical inputs, not derived from first principles.
  • Termination thresholds = e.g., wheel contact force > 350 N, rear wheel velocity > 77 rad/s (Table S9)
    Hand-selected safety limits; they affect how the policy balances risk and behavior, and are not derived from an external standard.
  • Domain randomization ranges = e.g., friction U[0.6,1.6], added mass scale U[0.9,1.1], heightmap drift U[-0.1,0.1] (Table S8)
    Chosen manually to span expected physical variation; these ranges are critical for successful sim-to-real transfer.
assumptions (5)
  • domain assumption The Isaac Lab dynamics and contact models, after manual wheel-contact tuning from drop tests, are sufficiently accurate for zero-shot policy transfer to hardware.
    Invoked throughout Methods (Sim-to-Real, Section S6). If the simulation were not faithful, the hardware successes would not occur, but the paper provides no formal sim-to-real gap analysis.
  • domain assumption Each policy in the orchestrator has a funnel or region of attraction that overlaps the entry set of the next policy, as required by sequential composition.
    Section S9 states the orchestrator is based on sequential composition [74] and that transitions are validated, but the funnels are not explicitly characterized or guaranteed.
  • domain assumption External motion capture and an offline terrain map provide accurate localization and heightmap inputs during hardware experiments.
    Section S8 states the platform lacks onboard exteroceptive sensors and the grid map is generated offline from known terrain; all autonomous waypoint and jump behaviors depend on this instrumentation.
  • domain assumption The reference trajectories used for motion imitation become dynamically feasible for UMV after iterative refinement via IMI.
    The flip and lateral-jump policies rely on IMI [65]; the method assumes that imperfect references can be iterated to feasible motions that the real robot can track.
  • ad hoc to paper Hand-designed reward functions and curriculum schedules are sufficient for PPO to discover the intended acrobatic behaviors.
    The reward tables (S3-S7) and curricula (S10) are constructed specifically for this platform and task set; there is no guarantee they generalize, and convergence to the desired stunt is an empirical outcome, not a theorem.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bicycle Acrobatics with Reinforcement Learning." pith.science (2026). https://pith.science/paper/GGTAT3WT

@misc{pith2026260800880,
  author       = {Pith},
  title        = {Pith review of: Bicycle Acrobatics with Reinforcement Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GGTAT3WT}},
  note         = {Machine review of arXiv:2608.00880}
}
read the original abstract

Bicycle robots are fast and energy efficient, but their simple mechanical design and their underactuated and non-holonomic dynamics make highly agile maneuvers difficult to achieve. Here, we use Reinforcement Learning (RL) to enable a bicycle robot to learn and compose a diverse repertoire of dynamic acrobatic stunts. Using different RL formulations such as waypoint following, pose reaching, twist tracking, guided tracking, and motion imitation, the robot acquires autonomous single and multi-table forward and lateral jumps, steerable jumps, front flips, kip-ups, kip-downs, driving, wheelies, bunny hops, and three-point turns. To coordinate these behaviors, we introduce an orchestrator that transitions between policies using state-dependent triggers, enabling robust long-horizon acrobatic stunts. We validate the approach on the Ultra Mobility Vehicle (UMV), a custom bicycle robot, in simulation and hardware. The robot repeatedly traverses tables up to 1 m high, performs more than 15 consecutive autonomous jumps while following waypoints, handles previously unseen multi-table configurations, executes continuous repertoires of kipups, jumps, flips, kip-downs, over more than 20 consecutive trials, and performs more than 10 consecutive autonomous and steerable repertoires of wheelies, lateral jumps, and single-wheel jump downs. These results demonstrate that RL can endow bicycle robots with levels of agility previously associated primarily with legged platforms while preserving the speed and efficiency of wheeled locomotion, establishing a foundation for bicycle acrobatics.

Figures

Figures reproduced from arXiv: 2608.00880 by the authors.

Figure 1
Figure 1. Ultra Mobility Vehicle (UMV). (A) The robot has eleven rigid links (green), five actuators (blue), and additional structural components (purple). The eleven links are the Head, left and right Fins, Neck, left and right Tie Rods, Bike, Fork, out￾of-plane Yoll, and front and rear Wheels. (B) The five actuated joints are shown in blue, and the passive joints are shown in red. The joints angles are the Head qh , Left Fi… view at source ↗
Figure 2
Figure 2. Orchestrated Kips, Jumps, & Flips. The robot’s trace (A) and corresponding key video frames (B–F) illustrate a single long orchestrated sequence of stunts on UMV. The stunts are performed seamlessly in the following order: (B) kip-up, (C) table jump up (D) table jump down, (E) front flip, (F) kip-down. This repertoire is robustly demonstrated across more than 20 consecutive trials [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗
Figure 3
Figure 3. Orchestrated Lateral Jumps & Wheelies. The robot’s trace and corresponding key video frames of autonomous (A) and steerable (B) wheelie and lateral jump stunts. The robot performs (B) solely by balancing on the rear wheel. Yellow arrows indicate the direction of motion [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Performance Analysis of the Kips, Jumps & Flips. (A) 3D trajectory of four consecutive loops. (B) Radar chart of key metrics across the four stunts. (C–H) Bar charts of the absolute peak values of these metrics across the stunts. The metrics are maximum power, regenera…
Figure 5
Figure 5. Figure 5: Autonomous & Steerable & Table Jumps Analysis. (A) 15 times autonomous 1 m table jump with a single policy. Green circles and red squares indicate start and end positions. Solid lines denote full trajectories. (B) 15 times 75 cm table jump with the same policy. (C) Top…
Figure 6
Figure 6. Figure 6: Autonomous & Steerable & Table Jumps Stunts. The figures present a collage of snapshots from the various experiments demonstrating the diverse set of jumping stunts and long repertoires achieved by UMV. In order of appearance from top left to bottom right, the figure i…
Figure 7
Figure 7. Figure 7: Additional Orchestrated Stunts. (A,B) Snapshots (i–iii) and the corresponding X–Z body trajectory of the kip up (A) and kip down (B). Overlaid snapshots of the (C) kip up and down, (D) front flip, (E) Three-point turn, and (F) bunny hop. we enabled the robot to seamles…
Figure 8
Figure 8. Figure 8: Autonomous execution of the combined kip, jump and flip stunt coordinated by the orchestrator. (A) A single typical hardware experiment of this stunt showing a kip-up (orange), jump/waypoint-following (blue), flip (green), and kip-down (pur￾ple). Numbered circles denot…

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

94 extracted references · 61 canonical work pages

  1. [1]

    Bokser,et al., System Design of the Ultra Mobility Vehicle: A Driving, Balancing, and Jumping Bicycle Robot.arXiv preprint arXiv:2602.22118(2026)

    B. Bokser,et al., System Design of the Ultra Mobility Vehicle: A Driving, Balancing, and Jumping Bicycle Robot.arXiv preprint arXiv:2602.22118(2026)

    arXiv 2026

  2. [2]

    Dynamics, Introducing Handle, Y outube video, https://www.youtube

    B. Dynamics, Introducing Handle, Y outube video, https://www.youtube. com/watch?v=-7xvqQeoA8c (2017), accessed: 15 September 2025

    2017

  3. [3]

    Kashiri,et al., Centauro: A hybrid locomotion and high power resilient manipulation platform.IEEE Robotics and Automation Letters 4(2), 1595–1602 (2019)

    N. Kashiri,et al., Centauro: A hybrid locomotion and high power resilient manipulation platform.IEEE Robotics and Automation Letters 4(2), 1595–1602 (2019)

    2019

  4. [4]

    Klemm,et al., Ascento: A two-wheeled jumping robot, in2019 International conference on robotics and automation (ICRA)(IEEE) (2019), pp

    V. Klemm,et al., Ascento: A two-wheeled jumping robot, in2019 International conference on robotics and automation (ICRA)(IEEE) (2019), pp. 7515–7521

    2019

  5. [5]

    Bjelonic, P

    M. Bjelonic, P . K. Sankar, C. D. Bellicoso, H. Vallery, M. Hutter, Rolling in the deep–hybrid locomotion for wheeled-legged robots using online Research Article - Preprint Robotics and AI Institute 14 kipup trigger Kip Up 0.8 sec Kip DownStart/End kipdown trigger0.8 sec 0.8 sec & drive trigger 0.8 sec & kipdown trigger 0.8 sec & drive trigger Fig. 8.Auto...

    2020

  6. [6]

    Bjelonic, V

    M. Bjelonic, V. Klemm, J. Lee, M. Hutter, A survey of wheeled-legged robots, inClimbing and walking robots conference(Springer) (2022), pp. 83–94

    2022

  7. [7]

    D. Liu, F . Y ang, X. Liao, X. Lyu, Diablo: A 6-dof wheeled bipedal robot composed entirely of direct-drive joints, in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)(IEEE) (2024), pp. 3605–3612

    2024

  8. [8]

    Unitree Robotics, Unitree B2-W (2025), https://www.unitree.com/b2-w

    2025

Show all 94 references
  1. [9]

    D. E. Jones, From the archives: The stability of the bicycle.Physics Today59(9), 51–56 (2006)

    2006

  2. [10]

    Astrom, R

    K. Astrom, R. Klein, A. Lennartsson, Bicycle dynamics and control: adapted bicycles for education and research.IEEE Control Systems Magazine25(4), 26–47 (2005),

    2005

  3. [11]

    J. P . Meijaard, J. M. Papadopoulos, A. Ruina, A. L. Schwab, Linearized dynamics equations for the balance and steer of a bicycle: a benchmark and review.Proceedings of the Royal society A: mathematical, physical and engineering sciences463(2084), 1955–1982 (2007)

    2007

  4. [12]

    J. D. G. Kooijman, J. P . Meijaard, J. M. Papadopoulos, A. Ruina, A. L. Schwab, A Bicycle Can Be Self-Stable Without Gyroscopic or Caster Effects.Science332(6027), 339–342 (2011),

    2011

  5. [13]

    L. Cui,et al., Nonlinear balance control of an unmanned bicycle: De- sign and experiments, in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)(IEEE) (2020), pp. 7279–7284

    2020

  6. [15]

    Xiong, R

    J. Xiong, R. Yu, C. Liu, Steering control and stability analysis for an autonomous bicycle: part II experiments and a modified linear control law.Nonlinear Dynamics112(5), 3107–3132 (2024)

    2024

  7. [16]

    B. Wang, F . Jing, Y . Deng, Z. Chen, B. Liang, Bayesian Optimization- Based Ideal Landing Planning for Ramp Jump of Single-Track Two- Wheeled Robots, in2024 IEEE International Conference on Robotics and Biomimetics (ROBIO)(IEEE) (2024), pp. 396–401

    2024

  8. [17]

    B. Wang, F . Jing, Y . Deng, Z. Chen, B. Liang, Attitude Control During the Flight Phase of Ramp Jump for Single-Track Two-Wheeled Robots, in2024 IEEE International Conference on Robotics and Biomimetics (ROBIO)(IEEE) (2024), pp. 1629–1634

    2024

  9. [18]

    Randløv, P

    J. Randløv, P . Alstrøm, Learning to Drive a Bicycle Using Reinforcement Learning and Shaping., inICML, vol. 98 (1998), pp. 463–471

    1998

  10. [19]

    X. Zhu,et al., Deep Reinforcement Learning-Based Control of Bicycle Robots on Rough Terrain, in2023 9th International Conference on Control, Automation and Robotics (ICCAR)(IEEE) (2023), pp. 103– 108

    2023

  11. [20]

    Zheng,et al., Reinforcement learning-based control of single-track two-wheeled robots in narrow terrain.Actuators12(3), 109 (2023)

    Q. Zheng,et al., Reinforcement learning-based control of single-track two-wheeled robots in narrow terrain.Actuators12(3), 109 (2023)

    2023

  12. [21]

    Baltes, G

    J. Baltes, G. Christmann, S. Saeedvand, A deep reinforcement learning algorithm to control a two-wheeled scooter with a humanoid robot. Engineering Applications of Artificial Intelligence126, 106941 (2023)

    2023

  13. [22]

    Zheng, D

    Q. Zheng, D. Wang, Z. Chen, Y . Sun, B. Liang, Continuous reinforce- ment learning based ramp jump control for single-track two-wheeled robots.Transactions of the Institute of Measurement and Control44(4), 892–904 (2022)

    2022

  14. [23]

    Z. Yuan, L. Y e, H. Liu, Z. Chen, B. Liang, Dynamic Jumps of an Agile Bicycle Through Reinforcement Learning, in2024 IEEE 3rd Industrial Electronics Society Annual On-Line Conference (ONCON) (IEEE) (2024), pp. 1–5

    2024

  15. [24]

    J. Tan, Y . Gu, C. K. Liu, G. Turk, Learning bicycle stunts.ACM Transac- tions on Graphics (TOG)33(4), 1–12 (2014)

    2014

  16. [25]

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

    1986

  17. [26]

    Kuindersma,et al., Optimization-based locomotion planning, esti- mation, and control design for the atlas humanoid robot.Autonomous Robots40(3), 429–455 (2016),

    S. Kuindersma,et al., Optimization-based locomotion planning, esti- mation, and control design for the atlas humanoid robot.Autonomous Robots40(3), 429–455 (2016),

    2016

  18. [27]

    Di Carlo, P

    J. Di Carlo, P . M. Wensing, B. Katz, G. Bledt, S. Kim, Dynamic Locomo- tion in the MIT Cheetah 3 Through Convex Model-Predictive Control, in2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)(2018), pp. 1–9,

    2018

  19. [28]

    C. D. Bellicoso, F . Jenelten, C. Gehring, M. Hutter, Dynamic Locomotion Through Online Nonlinear Motion Optimization for Quadrupedal Robots. IEEE Robotics and Automation Letters3(3), 2261–2268 (2018),

    2018

  20. [29]

    Fahmi, C

    S. Fahmi, C. Mastalli, M. Focchi, C. Semini, Passive Whole-Body Control for Quadruped Robots: Experimental Validation Over Challeng- ing Terrain.IEEE Robotics and Automation Letters4(3), 2553–2560 (2019),

    2019

  21. [30]

    Rudin, D

    N. Rudin, D. Hoeller, P . Reist, M. Hutter, Learning to Walk in Minutes Us- ing Massively Parallel Deep Reinforcement Learning, inProceedings of the 5th Conference on Robot Learning, A. Faust, D. Hsu, G. Neumann, Eds. (PMLR), vol. 164 ofProceedings of Machine Learning Research...

    2022

  22. [31]

    Ha,et al., Learning-based legged locomotion: State of the art and future perspectives.Int

    S. Ha,et al., Learning-based legged locomotion: State of the art and future perspectives.Int. J. Robot. Res.44(8), 1396–1427 (2025),

    2025

  23. [32]

    J. Lee, J. Hwangbo, L. Wellhausen, V. Koltun, M. Hutter, Learning quadrupedal locomotion over challenging terrain.Science robotics 5(47), eabc5986 (2020)

    2020

  24. [33]

    T. Miki,et al., Learning robust perceptive locomotion for quadrupedal Research Article - Preprint Robotics and AI Institute 15 robots in the wild.Science Robotics7(62), eabk2822 (2022), , https: //www.science.org/doi/abs/10.1126/scirobotics.abk2822

    2022 doi

  25. [34]

    Hoeller, N

    D. Hoeller, N. Rudin, D. Sako, M. Hutter, Anymal parkour: Learn- ing agile navigation for quadrupedal robots.Science Robotics9(88), eadi7566 (2024)

    2024

  26. [35]

    Cheng, K

    X. Cheng, K. Shi, A. Agarwal, D. Pathak, Extreme parkour with legged robots, in2024 IEEE International Conference on Robotics and Au- tomation (ICRA)(IEEE) (2024), pp. 11443–11450

    2024

  27. [36]

    Zhuang,et al., Robot Parkour Learning, inConference on Robot Learning(PMLR) (2023), pp

    Z. Zhuang,et al., Robot Parkour Learning, inConference on Robot Learning(PMLR) (2023), pp. 73–92

    2023

  28. [37]

    Li,et al., Robust and Versatile Bipedal Jumping Control through Re- inforcement Learning, inRobotics science and systems(RSS) (2023)

    Z. Li,et al., Robust and Versatile Bipedal Jumping Control through Re- inforcement Learning, inRobotics science and systems(RSS) (2023)

    2023

  29. [38]

    Lee,et al., Learning robust autonomous navigation and locomotion for wheeled-legged robots.Science Robotics9(89), eadi9641 (2024)

    J. Lee,et al., Learning robust autonomous navigation and locomotion for wheeled-legged robots.Science Robotics9(89), eadi9641 (2024)

    2024

  30. [39]

    Kim,et al., High-speed control and navigation for quadrupedal robots on complex and discrete terrain.Science Robotics10(102), eads6192 (2025)

    H. Kim,et al., High-speed control and navigation for quadrupedal robots on complex and discrete terrain.Science Robotics10(102), eads6192 (2025)

    2025

  31. [40]

    He,et al., Attention-based map encoding for learning generalized legged locomotion.Science Robotics10(105), eadv3604 (2025)

    J. He,et al., Attention-based map encoding for learning generalized legged locomotion.Science Robotics10(105), eadv3604 (2025)

    2025

  32. [41]

    Bussola, M

    R. Bussola, M. Focchi, G. Turrisi, C. Semini, L. Palopoli, Guided rein- forcement learning for omnidirectional 3d jumping in quadruped robots. arXiv preprint arXiv:2507.16481(2025)

    2025 arXiv

  33. [42]

    X. B. Peng, P . Abbeel, S. Levine, M. Van de Panne, Deepmimic: Example-guided deep reinforcement learning of physics-based charac- ter skills.ACM Transactions On Graphics (TOG)37(4), 1–14 (2018)

    2018

  34. [43]

    X. B. Peng,et al., Learning agile robotic locomotion skills by imitating animals, inRobotics: Science and Systems (RSS)(2020)

    2020

  35. [44]

    X. B. Peng, Z. Ma, P . Abbeel, S. Levine, A. Kanazawa, Amp: Adver- sarial motion priors for stylized physics-based character control.ACM Transactions on Graphics (ToG)40(4), 1–20 (2021)

    2021

  36. [45]

    X. B. Peng, Y . Guo, L. Halper, S. Levine, S. Fidler, Ase: Large-scale reusable adversarial skill embeddings for physically simulated charac- ters.ACM Transactions On Graphics (TOG)41(4), 1–17 (2022)

    2022

  37. [46]

    Tessler, Y

    C. Tessler, Y . Guo, O. Nabati, G. Chechik, X. B. Peng, Maskedmimic: Unified physics-based character control through masked motion in- painting.ACM Transactions on Graphics (TOG)43(6), 1–21 (2024)

    2024

  38. [47]

    G. B. Margolis, M. Wang, N. Fey, P . Agrawal, SoftMimic: Learn- ing Compliant Whole-body Control from Examples.arXiv preprint arXiv:2510.17792(2025)

    2025

  39. [48]

    Kang,et al., Learning Steerable Imitation Controllers from Unstruc- tured Animal Motions.arXiv preprint arXiv:2507.00677(2025)

    D. Kang,et al., Learning Steerable Imitation Controllers from Unstruc- tured Animal Motions.arXiv preprint arXiv:2507.00677(2025)

    2025

  40. [49]

    X. B. Peng, A. Kanazawa, J. Malik, P . Abbeel, S. Levine, Sfv: Rein- forcement learning of physical skills from videos.ACM Transactions On Graphics (TOG)37(6), 1–14 (2018)

    2018

  41. [50]

    Miller, S

    A. Miller, S. Fahmi, M. Chignoli, S. Kim, Reinforcement learning for legged robots: Motion imitation from model-based optimal control. arXiv preprint arXiv:2305.10989(2023)

    2023 arXiv

  42. [51]

    D. Kang, J. Cheng, M. Zamora, F . Zargarbashi, S. Coros, Rl+ model- based control: Using on-demand optimal control to learn versatile legged locomotion.IEEE Robotics and Automation Letters8(10), 6619–6626 (2023)

    2023

  43. [52]

    Y oum,et al., Imitating and finetuning model predictive control for robust and symmetric quadrupedal locomotion.IEEE Robotics and Automation Letters8(11), 7799–7806 (2023)

    D. Y oum,et al., Imitating and finetuning model predictive control for robust and symmetric quadrupedal locomotion.IEEE Robotics and Automation Letters8(11), 7799–7806 (2023)

    2023

  44. [53]

    Fuchioka, Z

    Y . Fuchioka, Z. Xie, M. Van de Panne, Opt-mimic: Imitation of opti- mized trajectories for dynamic quadruped behaviors.arXiv preprint arXiv:2210.01247(2022)

    2022 arXiv

  45. [54]

    Liu,et al., Opt2skill: Imitating dynamically-feasible whole-body trajectories for versatile humanoid loco-manipulation.arXiv preprint arXiv:2409.20514(2024)

    F . Liu,et al., Opt2skill: Imitating dynamically-feasible whole-body trajectories for versatile humanoid loco-manipulation.arXiv preprint arXiv:2409.20514(2024)

    2024

  46. [55]

    Y ang,et al., OmniRetarget: Interaction-Preserving Data Generation for Humanoid Whole-Body Loco-Manipulation and Scene Interaction, inarxiv preprint arXiv:2509.26633(2026), pp

    L. Y ang,et al., OmniRetarget: Interaction-Preserving Data Generation for Humanoid Whole-Body Loco-Manipulation and Scene Interaction, inarxiv preprint arXiv:2509.26633(2026), pp. 1–12

    2026 arXiv

  47. [56]

    Mahmood, N

    N. Mahmood, N. Ghorbani, N. F . Troje, G. Pons-Moll, M. J. Black, AMASS: Archive of motion capture as surface shapes, inProceedings of the IEEE/CVF international conference on computer vision(2019), pp. 5442–5451

    2019

  48. [57]

    F . G. Harvey, M. Yurick, D. Nowrouzezahrai, C. Pal, Robust motion in-betweening.ACM Transactions on Graphics (TOG)39(4), 60–1 (2020)

    2020

  49. [58]

    He,et al., OmniH2O: Universal and Dexterous Human-to- Humanoid Whole-Body Teleoperation and Learning.arXiv preprint arXiv:2406.08858(2024)

    T. He,et al., OmniH2O: Universal and Dexterous Human-to- Humanoid Whole-Body Teleoperation and Learning.arXiv preprint arXiv:2406.08858(2024)

    2024 arXiv

  50. [59]

    He,et al., Asap: Aligning simulation and real-world physics for learn- ing agile humanoid whole-body skills.arXiv preprint arXiv:2502.01143 (2025)

    T. He,et al., Asap: Aligning simulation and real-world physics for learn- ing agile humanoid whole-body skills.arXiv preprint arXiv:2502.01143 (2025)

    2025 arXiv

  51. [60]

    Chen,et al., GMT: General Motion Tracking for Humanoid Whole- Body Control.arXiv preprint arXiv:2506.14770(2025)

    Z. Chen,et al., GMT: General Motion Tracking for Humanoid Whole- Body Control.arXiv preprint arXiv:2506.14770(2025)

    2025 arXiv

  52. [61]

    Wu,et al., Perceptive humanoid parkour: Chaining dynamic human skills via motion matching.arXiv preprint arXiv:2602.15827(2026)

    Z. Wu,et al., Perceptive humanoid parkour: Chaining dynamic human skills via motion matching.arXiv preprint arXiv:2602.15827(2026)

    2026 arXiv

  53. [62]

    Sleiman, ZEST: Zero-shot Embodied Skill Transfer for Athletic Robot Control.arXiv preprint arXiv:2602.00401(2026)

    J.-P . Sleiman, ZEST: Zero-shot Embodied Skill Transfer for Athletic Robot Control.arXiv preprint arXiv:2602.00401(2026)

    2026

  54. [63]

    D. S. Brown, W. Goo, S. Niekum, Extrapolating beyond suboptimal demonstrations via inverse reinforcement learning from observations, inInternational Conference on Machine Learning (ICML)(2019), pp. 783–792

    2019

  55. [64]

    Y .-H. Wu, N. Charoenphakdee, H. Bao, V. Tangkaratt, M. Sugiyama, Imitation learning from imperfect demonstration, inInternational Con- ference on Machine Learning (ICML)(2019), pp. 6818–6827

    2019

  56. [65]

    J. Kim, S. Fahmi, S. Rho, S. Ha, G. Nelson, Flip Stunts on Bicycle Robots using Iterative Motion Imitation, inProc. IEEE Int. Conf. Robot. Automat. (ICRA)(Vienna, Austria) (2026), pp. 1–8

    2026

  57. [66]

    Li,et al., Learning agile skills via adversarial imitation of rough partial demonstrations, inConference on Robot Learning(PMLR) (2023), pp

    C. Li,et al., Learning agile skills via adversarial imitation of rough partial demonstrations, inConference on Robot Learning(PMLR) (2023), pp. 342–352

    2023

  58. [67]

    Zargarbashi, J

    F . Zargarbashi, J. Cheng, D. Kang, R. Sumner, S. Coros, RobotKeyfram- ing: Learning Locomotion with High-Level Objectives via Mixture of Dense and Sparse Rewards, inConference on Robot Learning(PMLR) (2025), pp. 916–932

    2025

  59. [68]

    Rho,et al., LineRides: Line-Guided Reinforcement Learning for Bicycle Robot Stunts.IEEE Robotics and Automation Letters(2026),

    S. Rho,et al., LineRides: Line-Guided Reinforcement Learning for Bicycle Robot Stunts.IEEE Robotics and Automation Letters(2026),

    2026

  60. [69]

    Y . Ma, F . Farshidian, M. Hutter, Learning Arm-Assisted Fall Damage Reduction and Recovery for Legged Mobile Manipulators, in2023 IEEE International Conference on Robotics and Automation (ICRA)(2023), pp. 12149–12155,

    2023

  61. [70]

    Huang,et al., Learning Humanoid Standing-up Control across Di- verse Postures.arXiv preprint arXiv:2502.08378(2025)

    T. Huang,et al., Learning Humanoid Standing-up Control across Di- verse Postures.arXiv preprint arXiv:2502.08378(2025)

    2025 arXiv

  62. [71]

    Strauch,et al., Robot Crash Course: Learning Soft and Stylized Falling.arXiv preprint arXiv:2511.10635(2025)

    P . Strauch,et al., Robot Crash Course: Learning Soft and Stylized Falling.arXiv preprint arXiv:2511.10635(2025)

    2025

  63. [72]

    MacAskill, Red Bull, Danny MacAskill’s Imaginate (2013), https: //www.youtube.com/watch?v=Sv3xVOs7_No, accessed: 2026-05-18

    D. MacAskill, Red Bull, Danny MacAskill’s Imaginate (2013), https: //www.youtube.com/watch?v=Sv3xVOs7_No, accessed: 2026-05-18

    2013

  64. [73]

    MacAskill, Redd Bull Bike, Danny MacAskill’s Gymnasium (2020), https://www.youtube.com/watch?v=fAEBNEscL0c, accessed: 2026-05- 18

    D. MacAskill, Redd Bull Bike, Danny MacAskill’s Gymnasium (2020), https://www.youtube.com/watch?v=fAEBNEscL0c, accessed: 2026-05- 18

    2020

  65. [74]

    R. R. Burridge, A. A. Rizzi, D. E. Koditschek, Sequential composition of dynamically dexterous robot behaviors.The International Journal of Robotics Research18(6), 534–555 (1999),

    1999

  66. [75]

    Rudin, D

    N. Rudin, D. Hoeller, M. Bjelonic, M. Hutter, Advanced Skills by Learn- ing Locomotion and Local Navigation End-to-End, inProc. IEEE/RSJ Int. Conf. Intell. Robot. Syst. (IROS)(2022), pp. 2497–2503,

    2022

  67. [76]

    Z. Xie, H. Y . Ling, N. H. Kim, M. Van De Panne, Allsteps: curriculum- driven learning of stepping stone skills, inComputer Graphics Forum (Wiley Online Library), vol. 39 (2020), pp. 213–224

    2020

  68. [77]

    Portelas, C

    R. Portelas, C. Colas, L. Weng, K. Hofmann, P .-Y . Oudeyer, Automatic curriculum learning for deep RL: A short survey, inInternational Joint Conference on Artificial Intelligence (IJCAI)(2020)

    2020

  69. [78]

    Hwangbo,et al., Learning agile and dynamic motor skills for legged robots.Science Robotics4(26), eaau5872 (2019)

    J. Hwangbo,et al., Learning agile and dynamic motor skills for legged robots.Science Robotics4(26), eaau5872 (2019)

    2019

  70. [79]

    E. Chane-Sane,et al., CaT: Constraints as Terminations for Legged Locomotion Reinforcement Learning, inIEEE/RSJ International Con- ference on Intelligent Robots and Systems (IROS)(2024)

    2024

  71. [80]

    Schulman, F

    J. Schulman, F . Wolski, P . Dhariwal, A. Radford, O. Klimov, Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347(2017)

    2017 arXiv

  72. [81]

    Mittal,et al., Isaac Lab: A GPU-Accelerated Simulation Frame- work for Multi-Modal Robot Learning.arXiv preprint arXiv:2511.04831 (2025)

    M. Mittal,et al., Isaac Lab: A GPU-Accelerated Simulation Frame- work for Multi-Modal Robot Learning.arXiv preprint arXiv:2511.04831 (2025). Research Article - Preprint Robotics and AI Institute 16

    2025 arXiv

  73. [82]

    Pinto, M

    L. Pinto, M. Andrychowicz, P . Welinder, W. Zaremba, P . Abbeel, Asymmetric actor critic for image-based robot learning.arXiv preprint arXiv:1710.06542(2017)

    2017 arXiv

  74. [83]

    Abdolhosseini, H

    F . Abdolhosseini, H. Y . Ling, Z. Xie, X. B. Peng, M. Van de Panne, On learning symmetric locomotion, inProceedings of the 12th ACM SIGGRAPH Conference on Motion, Interaction and Games(2019), pp. 1–10

    2019

  75. [84]

    J. Tobin,et al., Domain randomization for transferring deep neural net- works from simulation to the real world, in2017 IEEE/RSJ international conference on intelligent robots and systems (IROS)(IEEE) (2017), pp. 23–30

    2017

  76. [85]

    X. B. Peng, M. Andrychowicz, W. Zaremba, P . Abbeel, Sim-to-real transfer of robotic control with dynamics randomization, in2018 IEEE international conference on robotics and automation (ICRA)(IEEE) (2018), pp. 3803–3810

    2018

  77. [86]

    Tan,et al., Sim-to-real: Learning agile locomotion for quadruped robots.arXiv preprint arXiv:1804.10332(2018)

    J. Tan,et al., Sim-to-real: Learning agile locomotion for quadruped robots.arXiv preprint arXiv:1804.10332(2018)

    2018 arXiv

  78. [87]

    Choi,et al., Learning quadrupedal locomotion on deformable terrain

    S. Choi,et al., Learning quadrupedal locomotion on deformable terrain. Science Robotics8(74), eade2256 (2023),

    2023

  79. [88]

    Todorov, T

    E. Todorov, T. Erez, Y . Tassa, Mujoco: A physics engine for model- based control, in2012 IEEE/RSJ international conference on intelligent robots and systems(IEEE) (2012), pp. 5026–5033

    2012

  80. [89]

    Open Neural Network Exchange (ONNX), https://github.com/onnx/ onnx

  81. [90]

    Bellicoso, A

    D. Bellicoso, A. Wollschläger, Exploy: EXport and dePLOY Rein- forcement Learning policies (2026), https://github.com/rai-opensource/ exploy

    2026

  82. [91]

    UP Xtreme i12, https://up-board.org/up-xtreme-i12/?ADL-P-Board

  83. [92]

    Specialized Bicycle Components, 2022 Hotwalk Carbon, https://www.specialized.com/us/en/hotwalk-carbon/p/200221? color=322068-200221&searchText=94021-0005 (2022), accessed: 2026-06-11

    2022

  84. [93]

    Dellaert,et al., borglab/gtsam: Release 4.2.Zenodo(2023)

    F . Dellaert,et al., borglab/gtsam: Release 4.2.Zenodo(2023)

    2023

  85. [94]

    Goldfain,et al., Autorally: An Open Platform for Aggressive Au- tonomous Driving.IEEE Control Systems Magazine39(1), 26–55 (2019)

    B. Goldfain,et al., Autorally: An Open Platform for Aggressive Au- tonomous Driving.IEEE Control Systems Magazine39(1), 26–55 (2019)

    2019

  86. [95]

    bail out

    C. Forster, L. Carlone, F . Dellaert, D. Scaramuzza,IMU preintegration on Manifold for Efficient Visual-Inertial Maximum-a-Posteriori Estima- tion(2015). AcknowledgmentsWe thank the entire hardware, simulation, machine learning, and controls team of UMV including A. Pre- ston,...

    2015

Pith tools

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