Pith. sign in

REVIEW 4 major objections 3 minor 31 references

Multi-critic Learning for Whole-body End-effector Twist Tracking

T0 review · 4 major / 3 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A single reinforcement-learning policy can walk and track end-effector trajectories simultaneously when rewards are split across multiple critics and commands are given as 6D twists.

desk verdict Solid whole-body loco-manipulation paper with credible sub-5cm tracking and a useful reward-sensitivity result; the hand-set 3000-iteration curriculum switch needs an ablation before the recipe transfers. read the letter →

arxiv 2507.08656 v2 pith:2PEGRK3V submitted 2025-07-11 cs.RO

classification cs.RO
keywords loco-manipulationmulti-criticreinforcementlearningwhole-bodycontrolend-effectortwisttrackingquadrupedmanipulatorteacher-studentdistillationrewardsensitivity
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 is trying to establish that a quadrupedal robot with a manipulator arm can be controlled by one learned policy that walks and moves its end effector at the same time, without switching controllers or hand-tuning reward weights. Its proposed recipe has two parts: a multi-critic actor architecture that gives locomotion, manipulation, and foot-contact scheduling their own value functions, and a twist-based command representation that tells the policy the desired end-effector linear and angular velocity rather than only the target pose. In simulation and on hardware, the authors show that this policy tracks straight, circular, and semicircular end-effector trajectories more accurately than pose-based and hierarchical baselines, and that it is insensitive to reward scaling by factors of 5 and 10, while a single-critic policy fails to learn both tasks. If this holds, whole-body loco-manipulation becomes a more tractable RL problem because conflicting objectives no longer need to be balanced by a single scalar reward.

What carries the argument

The central mechanism is the multi-critic actor update combined with the twist command. Three critics estimate value functions for reward groups--locomotion, manipulation, and contact schedule--and the advantages from each critic are normalized and summed before the policy update, so no weighted average of rewards or manually tuned mixture coefficient is needed. The twist command is generated from the interpolated trajectory as $v_{\mathrm{EE}} = (r_i - r_{\mathrm{EE}})/\Delta t$ and $\omega_{\mathrm{EE}} = (\theta_i \ominus \theta_{\mathrm{EE}})/\Delta t$, given to the policy together with the final goal pose, so the policy receives explicit velocity targets instead of a rigid sequence of poses. A curriculum switches this twist command from the base frame to a gravity-aligned control frame after 3000 training iterations, which the authors find necessary because the control frame moves too much early in training for the state-reward pairs to be learned easily.

What would settle it

Train the same multi-critic policy on a robot with slower torso stabilization while keeping the 3000-iteration frame switch: if the end-effector tracking error spikes after the switch and never recovers, the fixed schedule is the weak point. Alternatively, run the full training with the twist command expressed only in the base frame and compare final tracking error; if it matches the curriculum result, the frame-switch assumption is unnecessary.

Watch

Extended reading notes

Core claim

The central claim is that a single RL policy can simultaneously walk and track end-effector trajectories, and that this becomes practical when reward signals are split into three groups each with its own critic--locomotion, manipulation, and foot-contact schedule--and when the tracking command is specified as a 6D twist (linear and angular velocity of the end effector) plus the final goal pose rather than as a sequence of poses. On these terms, the multi-critic policy achieves end-effector positional tracking errors of about 0.0358 m while walking and 0.0176 m while standing in the main simulation comparison, beats several pose-based and hierarchical baselines across linear, circular, and workspace-sweep trajectories, and tracks the tested trajectories on hardware at commanded speeds between 0.05 and 0.2 m/s. The paper also reports that separating the contact schedule into its own critic lets the policy generalize from the trained static walk to a trot at runtime, and that the multi-critic policy stays stable when locomotion or manipulation rewards are scaled by 5x or 10x, where a single-critic policy only learns one of the two objectives.

Load-bearing premise

The load-bearing premise is that switching the twist-command frame after a fixed 3000 training iterations works across robots, gaits, and reward scales, because the schedule is based on the observation that robots learn to walk with a stable torso within 3000 iterations rather than on a principled convergence criterion.

Editorial extensions

If this is right

  • A single policy can replace controller-switching pipelines for loco-manipulation, because one set of weights handles both walking and end-effector motion.
  • Twist-based task specification gives explicit velocity control, so trajectories of varying speed can be tracked without re-planning intermediate poses.
  • Separate critics make reward scaling within one group harmless, removing a class of tuning hyperparameters in multi-objective RL.
  • Because contact scheduling has its own critic, the policy generalizes from a static walk to a trot when the commanded foot swing pattern changes.
  • In the reported comparisons, the method tracks whole-body trajectories with lower position and velocity errors than pose-based, hierarchical, and whole-body MPC baselines.

Reading between the lines

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

  • Editorial inference: The 3000-iteration frame switch is an empirically fixed schedule, so it could become a learned or adaptive switch triggered by torso stability, an option the paper does not explore.
  • Editorial inference: The multi-critic decoupling is not specific to locomotion versus manipulation; the same reward-group splitting might stabilize training for other conflicting objectives, such as simultaneously tracking a camera target and obeying joint limits.
  • Editorial inference: Because the twist formulation directly commands velocity, downstream controllers could integrate it with impedance or force-control layers, a connection the paper does not make.
  • Editorial inference: A direct test of the velocity-aware claim would be to command trajectories at speeds far outside the trained 0.05-0.2 m/s range and measure whether error scales gently or sharply, since the paper only reports within-range speeds.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 3 minor

Summary. The paper proposes a reinforcement-learning framework for whole-body loco-manipulation on a quadruped with an arm. The method combines a multi-critic actor architecture, with separate critics for locomotion, manipulation, and foot-contact rewards, with a 6D twist-based end-effector command formulation. The policy is trained in a teacher-student setup in simulation and deployed on an ANYmal D with a Dynaarm. The paper reports simulation and hardware tracking results for linear, circular, and semicircular trajectories, comparisons against pose-based and MPC baselines, and ablation studies on command-frame curricula and critic architecture. The central claims are that the multi-critic policy can simultaneously walk and track end-effector trajectories with lower tracking errors than the compared controllers, and that it is robust to reward-weight scaling where a single-critic policy fails.

Significance. If the claims hold, the paper offers a useful recipe for whole-body loco-manipulation: decoupling reward signals via multiple critics and encoding velocity information directly in the task command appear to reduce reward-tuning burden and improve end-effector tracking during locomotion. The paper is strong in breadth of evaluation: it includes extensive simulation comparisons against multiple baselines, controlled experiments for trajectory types and command-frame curricula, and hardware demonstrations with video evidence. The multi-critic robustness result in Fig. 6 is a valuable empirical contribution. However, several load-bearing aspects are not yet supported: the frame-switch curriculum is an unablated hand-set constant, the baseline comparisons are not end-to-end, and the reported results lack variance estimates. These issues make the paper a good candidate for major revision rather than acceptance in its current form.

major comments (4)
  1. [§7.2.1, §5] The curriculum that switches the end-effector twist command from the base frame to the control frame is a load-bearing training schedule, but it is justified only by the observation that 'robots learn to walk with a stable torso within 3000 iterations' and is never ablated. All hardware results and all comparison tables are produced with this single constant, so if the switch time is not robust across robots, reward scales, or terrains, the reported tracking accuracy may not transfer. Moreover, §5 states that the pipeline works 'without requiring complex training curricula,' which is in tension with this frame-switch schedule and the terrain curriculum in §7.1.3. Please provide a sensitivity study of the switch iteration, or an adaptive criterion, and a more precise statement of which components are hand-set.
  2. [§4, Tables 2–3] The comparison with Ma and Taka in Tables 2 and 3 is not a comparison with those methods' whole-body controllers. Per §4, the leg actions come from the respective locomotion policy while the arm actions are produced by the proposed approach. This protocol tests the proposed arm controller on top of fixed locomotion policies, not the full controllers, so the conclusion that the approach outperforms existing methods is stronger than the evidence. Portela [8] is evaluated only while standing, so its rows do not test loco-manipulation either. For a fair test, either train the baseline whole-body methods on the same task or explicitly reframe the comparison as an arm-controller module comparison and soften the corresponding claims.
  3. [§4, Figs. 6 and 8, Table 1] Several central performance and robustness claims rest on single-run results without reported variance. Figure 6 and Figure 8 show RL training curves with no seeds or confidence intervals, and Table 1 reports hardware tracking errors from single runs. Since RL training is stochastic and hardware executions vary, claims such as 'multi-critic policy is robust to reward scaling' and the quantitative superiority in Table 1 need multiple seeds/runs with mean and standard deviation, or at least a statement of the number of runs and variability.
  4. [§3.3, §4, Table 3] The benefit of the twist-based formulation over pose-based tracking is not isolated as a variable. The comparison against Portela [8] in Table 3 changes the task representation, critic architecture, reward structure, and training details simultaneously. Therefore the lower errors cannot be attributed to the twist formulation alone. A controlled ablation that keeps all other components fixed and swaps pose-based and twist-based task representations is needed to support contribution (2), or the conclusion should be restricted to the joint framework.
minor comments (3)
  1. [Table 5] The command observation rows list 'Desired Feet swing heights' twice; one of these rows should presumably be 'Desired End-Effector final goal pose'.
  2. [Eq. (5) and Eq. (3)] The symbol θ is used for both foot phase offsets in Eq. (5) and orientation in Eq. (3); using distinct symbols would avoid confusion. Also, h_f is defined as a vector but h_max is never explicitly defined.
  3. [§7.1.2, Table 6] In the manipulation reward group, 'Arm action rate', 'Arm joint torque', and 'Arm joint velocity' use the subscript 'robot' rather than 'arm'; this is inconsistent with the symbol table in Table 7 and should be corrected.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the central claims are validated against external baselines and held-out trajectories; the twist command is a definitional rewrite of pose error, not a fitted prediction.

full rationale

The paper's main empirical claims are supported by comparisons against external methods (Ma, Taka 3-DoF/6-DoF, whole-body MPC, and Portela) on trajectories that are not part of the training distribution, with tracking errors measured in simulation and on hardware (Tables 1-3). No parameter is fitted to these target errors: the multi-critic architecture, reward weights, and curriculum are fixed before evaluation. The twist-based command in Eq. (4) does define the desired end-effector velocity as the current pose error divided by the control time step, so the manipulation reward in Table 6 is algebraically a one-step pose-tracking target when the commanded twist comes from this generator. That is a definitional identity in the command formulation, not a circular derivation of a prediction, and the paper additionally evaluates explicit velocity commands at different speeds (Table 1), giving the velocity-control claim independent empirical content. The hand-set curriculum switch at iteration 3000 in Sec. 7.2.1 is an unablated robustness concern, but it is not circular because the reported tracking accuracy is not used to set that constant and the claim does not rest on a self-citation. Self-citations, such as [8] for the robot-centric task frame and [16] for teacher-student distillation, support standard design choices rather than the central empirical result. Overall, the derivation chain is self-contained against external benchmarks, so no load-bearing circularity is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The multi-critic architecture and twist commands are methods, not invented entities. The free parameters are the hand-tuned reward weights, reward variances, and the curriculum switch iteration; the axioms are standard math plus domain assumptions about frame choice and trajectory interpolation, with one ad hoc scheduling rule.

free parameters (3)
  • Reward weights within each critic group = e.g., loco 2.0, mani 5.0/4.0, contact 1.0
    Hand-tuned weights in Table 6; the paper argues multi-critic reduces cross-group tuning but within-group weights are still manually chosen.
  • Curriculum switch iteration = 3000
    Switch from base frame to control frame for twist commands at iteration 3000, set based on the observation that policies learn to stabilize the torso within 3000 iterations (Appendix 7.2.1).
  • Gaussian reward variances = sigma values e.g., 0.005, 0.01, 0.1, 0.05
    Variances in the tracking rewards (Table 6) set the desired precision; they are chosen by hand, not learned.
assumptions (4)
  • standard math SO(3) interpolation uses the Lie group operations circled plus and circled minus (Eq. 3).
    Group operations for orientation interpolation are standard and unproved in the paper.
  • domain assumption Twist commands expressed in a robot-centric gravity-aligned control frame decouple the manipulation task from torso pitch and roll.
    Section 3.2 claims this decoupling helps learning; no theoretical guarantee is provided.
  • domain assumption Training trajectories are linear interpolations between start and goal poses, and curved test trajectories (circle, semicircle) will still be trackable.
    Section 3.2 generates waypoints by linear interpolation; Section 4 tests circular shapes outside this distribution and reports good tracking.
  • ad hoc to paper The curriculum switch from base frame to control frame at iteration 3000 is a valid schedule.
    Section 7.2.1 sets the switch based on an observed training milestone, not a principled rule.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-critic Learning for Whole-body End-effector Twist Tracking." pith.science (2026). https://pith.science/paper/2PEGRK3V

@misc{pith2026250708656,
  author       = {Pith},
  title        = {Pith review of: Multi-critic Learning for Whole-body End-effector Twist Tracking},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2PEGRK3V}},
  note         = {Machine review of arXiv:2507.08656}
}
read the original abstract

Learning whole-body control for locomotion and arm motions in a single policy has challenges, as the two tasks have conflicting goals. For instance, efficient locomotion typically favors a horizontal base orientation, while end-effector tracking may benefit from base tilting to extend reachability. Additionally, current Reinforcement Learning (RL) approaches using a pose-based task specification lack the ability to directly control the end-effector velocity, making smoothly executing trajectories very challenging. To address these limitations, we propose an RL-based framework that allows for dynamic, velocity-aware whole-body end-effector control. Our method introduces a multi-critic actor architecture that decouples the reward signals for locomotion and manipulation, simplifying reward tuning and allowing the policy to resolve task conflicts more effectively. Furthermore, we design a twist-based end-effector task formulation that can track both discrete poses and motion trajectories. We validate our approach through a set of simulation and hardware experiments using a quadruped robot equipped with a robotic arm. The resulting controller can simultaneously walk and move its end-effector and shows emergent whole-body behaviors, where the base assists the arm in extending the workspace, despite a lack of explicit formulations. Videos and supplementary material can be found at multi-critic-locomanipulation.github.io.

Figures

Figures reproduced from arXiv: 2507.08656 by the authors.

Figure 1
Figure 1. Smooth and precise whole-body loco-manipulation. We present a framework that enables quadrupedal robots to perform precise loco-manipulation through a multi-critic architec￾ture and twist-based task space control. Our approach generates coordinated whole-body behaviors (Left) while maintaining accurate end-effector control during locomotion (Right). Abstract: Learning whole-body control for locomotion and arm motion… view at source ↗
Figure 2
Figure 2. Architecture for the teacher training pipeline: Given a randomly sampled start and goal pose, the desired end-effector and base twist, along with the desired foot height, are provided as commands to the policy by the command generator. Rewards are categorically computed and consumed by separate critics, leading to individual value functions. The advantage is estimated per critic, normalized, and summed to compute th… view at source ↗
Figure 3
Figure 3. End-Effector Tracking Performance for Different Trajectories: Desired and measured position and velocity for tracking linear, circular, and semicircular trajectories (left to right). • Circle: The policy is commanded to move in a circle on the YZ plane with a radius of 20 cm at a fixed distance in front of the robot. • Semicircle around robot: To evaluate trajectories over an extended workspace, the policy is comman… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Contact Schedule Adaptation to Unseen Gait Patterns: Measured height of Left Fore (LF), Right Fore (RF), Left Hind (LH) and Right Hind (RH) feet for (a) static walking measured on hardware, and (b) trot gait measured in simulation. of gait patterns, allowing it to adap…
Figure 5
Figure 5. Figure 5: Workspace sweep trajectory wrt. base used for [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Reward sensitivity analysis of single-critic vs multi-critic frameworks: Multi-critic policy is robust against reward scaling across categories whereas the single-critic policy only learns to perform the objective with higher reward scale. Reward Sensitivity Analysis -…
Figure 7
Figure 7. Figure 7: , representing the twist command in the control frame makes it challenging for the policy to learn to precisely follow the end-effector trajectory. We observe that as the robot learns to stabi￾0 10000 20000 30000 40000 50000 Training Steps 0.1 0.2 0.3 0.4 0.5 0.6 0.7 E…
Figure 8
Figure 8. Figure 8: Comparison of single-critic vs multi-critic frameworks: Multi-critic policy learns to track both the base and end-effector commands whereas the single-critic policy only learns to track the end-effector commands. policy that adopts a stationary posture at all times tha…
Figure 9
Figure 9. Figure 9: End-Effector Tracking performance in chicken-head mode: Tracking error in end effector posi￾tion (red, left axis) and measured base velocity (blue, right axis) of the robot plotted over time. The position error has a mean of 0.0156 m and maximum of 0.0287 m. To demonst…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 28 canonical work pages

  1. [8]

    Portela, A

    T. Portela, A. Cramariuc, M. Mittal, and M. Hutter. Whole-body end-effector pose tracking. In IEEE International Conference on Robotics and Automation (ICRA) , 2025

  2. [1]

    H. Ha, Y . Gao, Z. Fu, J. Tan, and S. Song. UMI on Legs: Making Manipulation Policies Mobile with Manipulation-Centric Whole-body Controllers. In Conference on Robot Learning (CoRL), 2024

  3. [2]

    Sleiman, F

    J.-P. Sleiman, F. Farshidian, M. V . Minniti, and M. Hutter. A Unified MPC Framework for Whole-Body Dynamic Locomotion and Manipulation. IEEE Robotics and Automation Letters (RA-L), 6(3):4688–4695, 2021

  4. [3]

    Mittal, D

    M. Mittal, D. Hoeller, F. Farshidian, M. Hutter, and A. Garg. Articulated Object Interaction in Unknown Scenes with Whole-Body Mobile Manipulation. In IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS) , pages 1647–1654, 2022

  5. [4]

    Zimmermann, R

    S. Zimmermann, R. Poranne, and S. Coros. Go Fetch! - Dynamic Grasps using Boston Dy- namics Spot with External Robotic Arm. In IEEE International Conference on Robotics and Automation (ICRA), pages 4488–4494, 2021

  6. [5]

    Chiu, J.-P

    J.-R. Chiu, J.-P. Sleiman, M. Mittal, F. Farshidian, and M. Hutter. A Collision-Free MPC for Whole-Body Dynamic Locomotion and Manipulation, 2022

  7. [6]

    P. Arm, M. Mittal, H. Kolvenbach, and M. Hutter. Pedipulate: Enabling Manipulation Skills using a Quadruped Robot’s Leg. In IEEE International Conference on Robotics and Automa- tion (ICRA), pages 5717–5723, 2024

  8. [7]

    Z. Fu, X. Cheng, and D. Pathak. Deep Whole-Body Control: Learning a Unified Policy for Manipulation and Locomotion. In Conference on Robot Learning (CoRL) , pages 138–149, 2023

Show all 31 references
  1. [9]

    Z. Wang, Y . Jia, L. Shi, H. Wang, H. Zhao, X. Li, J. Zhou, J. Ma, and G. Zhou. Arm- Constrained Curriculum Learning for Loco-Manipulation of a Wheel-Legged Robot. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 10770– 10776, 2024. 9

  2. [10]

    Jiang, Z

    K. Jiang, Z. Fu, J. Guo, W. Zhang, and H. Chen. Learning Whole-Body Loco-Manipulation for Omni-Directional Task Space Pose Tracking With a Wheeled-Quadrupedal-Manipulator. IEEE Robotics and Automation Letters (RA-L) , 10(2):1481–1488, 2025

  3. [11]

    Portela, G

    T. Portela, G. B. Margolis, Y . Ji, and P. Agrawal. Learning Force Control for Legged Ma- nipulation. In IEEE International Conference on Robotics and Automation (ICRA) , pages 15366–15372, 2024

  4. [12]

    Y . Ma, F. Farshidian, T. Miki, J. Lee, and M. Hutter. Combining Learning-Based Locomotion Policy With Model-Based Manipulation for Legged Mobile Manipulators. IEEE Robotics and Automation Letters (RA-L), 7(2):2377–2384, 2022

  5. [13]

    M. Liu, Z. Chen, X. Cheng, Y . Ji, R. Yang, and X. Wang. Visual Whole-Body Control for Legged Loco-Manipulation. In Conference on Robot Learning (CoRL) , 2024

  6. [14]

    G. Pan, Q. Ben, Z. Yuan, G. Jiang, Y . Ji, S. Li, J. Pang, H. Liu, and H. Xu. RoboDuet: Whole- body Legged Loco-Manipulation with Cross-Embodiment Deployment, 2024

  7. [15]

    Hwangbo, J

    J. Hwangbo, J. Lee, A. Dosovitskiy, D. Bellicoso, V . Tsounis, V . Koltun, and M. Hutter. Learn- ing agile and dynamic motor skills for legged robots. Science Robotics, 4(26):eaau5872, 2019

  8. [16]

    J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter. Learning quadrupedal locomo- tion over challenging terrain. Science Robotics, 5(47):eabc5986, 2020

  9. [17]

    T. Miki, J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter. Learning robust per- ceptive locomotion for quadrupedal robots in the wild. Science Robotics, 7(62), 2022. ISSN 2470-9476

  10. [18]

    Hoeller, N

    D. Hoeller, N. Rudin, D. Sako, and M. Hutter. ANYmal parkour: Learning agile navigation for quadrupedal robots. Science Robotics, 9(88):eadi7566, 2024

  11. [19]

    Y . Zhou, C. Barnes, J. Lu, J. Yang, and H. Li. On the Continuity of Rotation Representations in Neural Networks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 5745–5753, 2019

  12. [20]

    Mysore, G

    S. Mysore, G. Cheng, Y . Zhao, K. Saenko, and M. Wu. Multi-Critic Actor Learning: Teaching RL Policies to Act with Style . In International Conference on Learning Representations (ICLR), 2022

  13. [21]

    Cheng, L

    G. Cheng, L. Dong, W. Cai, and C. Sun. Multi-Task Reinforcement Learning With Attention- Based Mixture of Experts. IEEE Robotics and Automation Letters (RA-L) , 8(6):3812–3819, 2023

  14. [22]

    Carron, E

    A. Carron, E. Arcari, M. Wermelinger, L. Hewing, M. Hutter, and M. N. Zeilinger. Data- driven model predictive control for trajectory tracking with a robotic arm. IEEE Robotics and Automation Letters (RA-L), 4(4):3758–3765, 2019

  15. [23]

    Ferrolho, W

    H. Ferrolho, W. Merkt, V . Ivan, W. Wolfslag, and S. Vijayakumar. Optimizing Dynamic Trajec- tories for Robustness to Disturbances Using Polytopic Projections. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 7477–7484, 2020

  16. [24]

    Ferrolho, V

    H. Ferrolho, V . Ivan, W. Merkt, I. Havoutis, and S. Vijayakumar. RoLoMa: Robust loco- manipulation for quadruped robots with arms. Autonomous Robots, 47(8):1463–1481, 2023

  17. [25]

    Rudin, D

    N. Rudin, D. Hoeller, P. Reist, and M. Hutter. Learning to Walk in Minutes Using Massively Parallel Deep Reinforcement Learning. In Conference on Robot Learning (CoRL) , pages 91– 100, 2022. 10

  18. [26]

    Mittal, C

    M. Mittal, C. Yu, Q. Yu, J. Liu, N. Rudin, D. Hoeller, J. L. Yuan, R. Singh, Y . Guo, H. Mazhar, A. Mandlekar, B. Babich, G. State, M. Hutter, and A. Garg. Orbit: A Unified Simulation Framework for Interactive Robot Learning Environments. IEEE Robotics and Automation Letters (...

  19. [27]

    Zargarbashi, J

    F. Zargarbashi, J. Cheng, D. Kang, R. Sumner, and S. Coros. RobotKeyframing: Learning Locomotion with High-Level Objectives via Mixture of Dense and Sparse Rewards. arXiv preprint arXiv:2407.11562, 2024

  20. [28]

    J. Lee, L. Schroth, V . Klemm, M. Bjelonic, A. Reske, and M. Hutter. Exploring Constrained Reinforcement Learning Algorithms for Quadrupedal Locomotion. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 11132–11138, 2024

  21. [29]

    Hutter, C

    M. Hutter, C. Gehring, D. Jud, A. Lauber, C. D. Bellicoso, V . Tsounis, J. Hwangbo, K. Bodie, P. Fankhauser, M. Bloesch, et al. Anymal-a highly mobile and dynamic quadrupedal robot. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 38–44. IEEE, 2016

  22. [30]

    T. Miki, J. Lee, L. Wellhausen, and M. Hutter. Learning to walk in confined spaces using 3d representation. In IEEE International Conference on Robotics and Automation (ICRA) , pages 8649–8656, 2024. doi:10.1109/ICRA57147.2024.10610271

  23. [31]

    Schulman, F

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 11 7 Appendix 7.1 Training Details 7.1.1 Observation Space The observation vector st ∈ R187 consists of five main components: joint p...

Pith tools

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