Efficient Learning of A Unified Policy For Whole-body Manipulation and Locomotion Skills

Chengrui Zhu; Chuang Guo; Dianyong Hou; Yong Liu; Zhen Zhang; Zhibin Li

REVIEW 4 major objections 7 minor 26 references

Efficient Learning of A Unified Policy For Whole-body Manipulation and Locomotion Skills

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A physical feasibility-guided reward, computed from the arm's kinematics, leads one reinforcement-learning policy to learn whole-body loco-manipulation without getting stuck in locomotion-first local optima.

desk verdict Useful reward-shaping idea with a real circularity problem in the headline metric; worth engaging for the hardware demos, but the workspace-expansion number is not independent evidence. read the letter →

arxiv 2507.04229 v1 pith:ILYIZKUJ submitted 2025-07-06 cs.RO cs.SYeess.SY

Dianyong Hou , Chengrui Zhu , Zhen Zhang , Zhibin Li , Chuang Guo , Yong Liu This is my paper · ORCID

classification cs.ROcs.SYeess.SY MSC 68T40

keywords whole-bodycontrolleggedmanipulatorreinforcementlearningkinematicmodelphysicalfeasibility-guidedrewardloco-manipulationsim-to-realtransferquadrupedrobot

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that reinforcement learning for a quadruped robot with a manipulator gets stuck in locomotion-first local optima, and that this can be avoided by feeding the policy an explicit kinematic model of the arm as a reward signal. The physical feasibility-guided (PFG) reward pays the agent whenever the torso posture makes the target end-effector pose reachable by the arm alone, so the policy explores whole-body postures early instead of dropping them for velocity tracking. The claim is that this yields a single unified policy with higher end-effector accuracy than other deep RL methods, expands the arm's usable workspace by about 34% while keeping velocity tracking degradation at or below 3%, and deploys zero-shot on a real quadruped. A sympathetic reader would take the paper's central contribution to be a parameter-light way to inject geometry into the reward, making difficult loco-manipulation skills learnable in about nine hours of simulated training.

What carries the argument

The load-bearing object is the physical feasibility-guided reward, defined by the feasible-state function $FS(s_{\text{torso}}, {}^W T)$, which is 1 exactly when an inverse-kinematics solver finds an arm joint configuration that places the end-effector at the target pose in the body frame. The paper strips the full-body model out of this check by expressing the target in the torso-dependent body frame, so only the arm's product-of-exponentials forward kinematics and a damped least-squares IK solver are needed. That boolean feasibility signal, added as a positive reward term $r_{\text{feasible}}$ with weight 0.16, is what biases exploration toward coordinated torso postures early in training. A cubic-polynomial command planner generates smooth reference trajectories in a body-projected coordinate system, which the paper argues reduces the learning burden of pose tracking while keeping trajectories within range as the robot moves.

What would settle it

Train the identical policy with the feasibility oracle changed from arm-only IK to a whole-body check that includes leg joint limits, contact constraints, and static balance for the commanded torso pose, keeping all reward weights and the command planner fixed; if the IK solution rate and 60th-percentile pose error do not worsen, the paper's claim that the arm-only kinematic model drives the improvement is falsified.

Watch

Extended reading notes

Core claim

The central discovery is that the local-optima problem in whole-body loco-manipulation is, to a large degree, an exploration problem, and that a purely kinematic feasibility oracle can fix it. At each step the method checks whether, for the current torso pose, there exists an arm configuration $q_{\text{arm}} \in \Theta_{\text{arm}}$ such that the arm's forward kinematics reaches the target expressed in the torso frame, $FK_{\text{arm}}(q_{\text{arm}}) = {}^B T(s_{\text{torso}})$. The boolean $FS(s_{\text{torso}}, {}^W T)$ becomes a positive reward term, so the policy is rewarded for adopting torso postures that enlarge the arm's reach even before the arm can actually track the target. The paper reports that this reward keeps the inverse-kinematics solution rate stable during the phase when velocity-tracking rewards are growing, whereas the baselines see their IK solution rate drop from 43% to 37% in that phase. In the final evaluation the method reaches a 60th-percentile end-effector position error of 8.7 cm and rotation error of 0.18 rad, compared with 13.1-14.4 cm for the baselines, and the learned policy transfers zero-shot to the physical robot.

Load-bearing premise

The method assumes that an arm-only inverse-kinematics check on the body-relative target pose is a correct oracle for what the whole robot, including its legs, balance, and contacts, can actually reach.

Editorial extensions

If this is right

A single actor network can output all 18 target joint positions for body and arm together, so no separate locomotion and manipulation controllers are needed for the demonstrated tasks.
The PFG reward transfers across robots with different arm kinematics because it relies only on the arm's product-of-exponentials model, not on a hand-built whole-body model.
Because the feasibility reward is always positive when the torso is useful, the policy keeps whole-body cooperative postures during the locomotion-learning phase, which is the mechanism the paper gives for avoiding local optima.
The 34% workspace expansion, measured by IK solution rate, implies that substantially more end-effector targets are reachable through body cooperation than through the arm alone.
The zero-shot hardware deployment supports the claim that the learned behavior, including adaptive body pitch during vertical arm movements, survives the sim-to-real gap under domain randomization.

Reading between the lines

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

An implicit consequence of using an arm-only feasibility oracle is that the reward can be maximized by torso postures that the legs cannot actually hold; upgrading $FS$ to include leg joint limits, contact modes, and static balance would test whether the kinematic-only check is a lower or upper bound on the reported gains.
The paper does not analyze how the always-positive feasibility reward distorts the value baseline; in principle it inflates the critic's value estimate in feasible regions, which could change exploration variance in ways unrelated to the geometric guidance.
A natural distributional test would re-evaluate the 8.7 cm error on targets sampled densely near the ground and behind the torso, where the paper itself notes accuracy degrades; the headline percentile is defined over targets sampled from the arm's forward kinematics, so it is not a workspace-wide guarantee.
Since the command planner and the PFG reward are introduced together, re-running the ablation with the planner removed, or with random waypoints, would isolate how much of the accuracy gain comes from smooth reference trajectories rather than from the feasibility signal.

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.

Desk Editor's Note

read the letter

The paper's genuine contribution is the physical feasibility-guided reward: a damped least-squares IK oracle that scores whether the torso posture makes a manipulator target reachable, then feeds that into PPO as an extra reward. That is a clean, easy-to-understand idea, and the authors integrate it carefully with a POE kinematic model and a body-projected command planner. The real-robot demonstrations on the X20-Z1—bending front legs to reach near-ground targets, pitching the torso to enlarge vertical reach, and the ribbon-waving stunt—are impressive and suggest the method produces coordinated behavior that plain baselines would not find. Credit is also due for reporting the velocity-tracking degradation explicitly and for discussing the non-uniform accuracy across the workspace.

The soft spot is real and it is the load-bearing quantitative claim. The IK solution rate used as the ablation metric is the same feasibility function that defines the PFG reward. Measuring that this rate rises is largely measuring reward optimization, not an independent outcome. The 34% workspace-expansion claim therefore overstates what the data show. The performance validation in Table III has a related issue: the 5000 evaluation targets are generated from the manipulator's forward kinematics, which by construction are reachable by the arm alone, so the improved PE/RE may reflect better arm tracking rather than expanded whole-body workspace. There are also no error bars or multi-seed statistics, and no code or data are provided, which makes it hard to push past these concerns. The hardware section is qualitative—compelling visually, but not a substitute for a quantitative task evaluation.

None of this is fatal. The idea is plausible, the reward-shaping mechanism is transparent, and the paper clearly shows that a hand-crafted kinematic feasibility signal can preserve cooperative torso postures during early locomotion learning. But the current manuscript does not establish the headline claim as stated. A serious referee would need to see an independent reachability metric, task-based evaluation on targets that require torso movement, and seed-to-seed variation.

This paper is for researchers working on legged manipulation and reward design. It deserves peer review and revision, not a desk reject. If I were editing, I would send it out with a clear request to fix the evaluation methodology.

Referee Report

4 major / 7 minor

Summary. The paper proposes Physical Feasibility-Guided (PFG) reward shaping for training a unified reinforcement-learning policy for a quadruped robot with a manipulator. The key idea is to use an off-the-shelf arm inverse-kinematics solver as an oracle: if the current torso pose makes the arm's IK succeed for a given target, the policy receives a positive reward. This is intended to prevent the policy from abandoning whole-body cooperative postures (e.g., crouching) during early locomotion-focused optimization. The authors validate in simulation with PPO, comparing AAC+PFG against AAC and ROA+AM baselines, and report that PFG raises the IK solution rate by 34%, improves end-effector position/rotation error at the 60th percentile (0.087 m / 0.18 rad), and deploys zero-shot on a DeepRobotics X20-Z1 robot, with hardware experiments showing coordinated pitch/height tracking and velocity tracking.

Significance. If the claims held, the contribution would be a simple, portable way to inject kinematic priors into model-free RL for legged manipulators, with broad applicability. The hardware demonstrations (ribbon waving, object pickup, cart pushing) are a genuine strength, and the paper is clearly written. However, the central quantitative evidence is weakened by a circular evaluation: the headline workspace metric is the same quantity that the PFG reward optimizes. The evaluation also relies on an arm-only kinematic oracle and on target poses generated from the manipulator's forward kinematics, so the experiments do not yet independently establish that PFG expands the true whole-body workspace or avoids local optima. With an independent metric and appropriate caveats, the method could be a useful contribution.

major comments (4)

[Section IV.A, Fig. 5, Eq. (2), Algorithm 1] The IK Solution Rate is the same feasibility condition used to define r_feasible in Table II. The PFG policy is rewarded exactly when FS(storso, W_T)=1, and the ablation metric reports the fraction of sampled targets for which the same damped-least-squares IK succeeds. The 34% workspace-expansion claim is therefore circular: it demonstrates that the policy increases the rewarded quantity, not that whole-body coordination improved. Please report an independent metric, such as the fraction of targets actually reached by the deployed end-effector within a tolerance, or a reachable-volume comparison computed with whole-body IK that includes leg joint limits and contact constraints.
[Section IV.B, Table III] The performance evaluation uses 5000 target poses 'randomly generated based on the forward kinematics of the manipulator.' Because these targets are arm-reachable by construction, the improved PE/RE of AAC+PFG may reflect better tracking of arm-reachable targets rather than a genuinely expanded whole-body workspace. In addition, PE and RE are reported as single 60th-percentile values without error bars or number of seeds, so the comparison is not statistically grounded. Please evaluate on targets that are not filtered by the arm-only oracle and report variance over seeds.
[Section III.B, Eq. (2)] The feasibility oracle eliminates the full-body model and checks only whether the arm IK can reach B_T(storso). It does not verify that the legs can realize storso or that the robot can maintain balance and contact. The motivating example in Section III.B (bending the front legs to reach a ground target) inherently depends on leg feasibility, so the PFG reward may systematically encourage torso postures that are not physically achievable. Please quantify the mismatch between the arm-only oracle and a whole-body reachability check, or at least discuss and bound this simplification.
[Section IV.A, Fig. 5] The claim that PFG 'prevents convergence to local optima' is not directly supported. The comparison adds a reward term to the objective, so the observed Phase B drop in IK success for AAC and ROA+AM can be explained by a different reward trade-off rather than by a local optimum. To support the local-optima claim, show that a baseline with the same task rewards and comparable exploration converges to a strictly worse scalar objective, or provide additional diagnostics (e.g., multiple seeds, success-rate on the original task reward).

minor comments (7)

[Throughout] Use 'AAC' consistently; Fig. 5 and some text use 'ACC'.
[Section IV.A] Define how the 34% workspace expansion is computed from the IK solution rate, and whether the reported percentage is the final gap or an integral over training.
[Section IV.B] Specify how the 5000 target poses are generated and whether they are filtered for arm reachability; the current wording is ambiguous.
[Table III] Report the number of evaluation episodes/rollouts and include confidence intervals for PE and RE.
[Algorithm 1, Table II] The notation is inconsistent ('ideal qt' vs 'ideal q(i)', 'kinematics max(...)'); please unify and define all symbols.
[Section IV.A] There are typos such as 'velocoty tracking rewrad' and 'ACC' for 'AAC'; please proofread.
[Title/Abstract] The title and abstract claim 'efficient learning,' but no wall-clock or sample-efficiency comparison with baselines is provided; consider adding such a comparison.

Circularity Check

2 steps flagged · score 6.0 of 10

The headline workspace-expansion metric is the PFG reward's own IK feasibility oracle, and the simulation test targets are arm-reachable by construction, so the central evidence partially reduces to the method's input.

self definitional [Section IV-A (Ablation Studies) vs Section III-B/C]
"1. Inverse Kinematics(IK) Solution Rate: The command planner samples the end-effector’s target in real-time and calculates the proportion of successful inverse kinematics solutions during the training process. This metric reflects legged manipulator’s ability to achieve cooperative motion."

The PFG reward is defined by FS(storso,W_T)=1 iff the same damped least-squares arm-IK solver (Algorithm 1) finds qarm with FK_arm(qarm)=B_T(storso); Section III-C says r_feasible is awarded if and only if FS=1. The ablation metric 'IK Solution Rate' is therefore the empirical frequency of exactly the event that triggers the PFG reward. Reporting that AAC+PFG keeps this rate high while baselines drop restates that the policy was rewarded for that event; the paper's 'workspace by 34% (based on the IK solution rate)' conclusion is not independent evidence of whole-body coordination.
self definitional [Section IV-B (Performance Validations)]
"In the simulation, 5000 target poses were randomly generated based on the forward kinematics of the manipulator."

Sampling targets from the manipulator's forward kinematics guarantees, by construction, that each target has a joint configuration qarm with FK_arm(qarm)=B_T, i.e. FS=1 in Eq. (2). The evaluation set is thus confined to the same arm-IK feasible set that the PFG reward and critic feature ('inverse kinematics solver status') are built on. Improved PE/RE on these arm-reachable targets cannot demonstrate that body-posture changes expanded the whole-body workspace; that claim rests entirely on the reward-identical IK solution rate.

full rationale

The central contribution is a reward-shaping term (PFG) built from an arm-IK feasibility oracle. The two main quantitative supports for the workspace-expansion claim are not independent of that oracle. First, the ablation metric 'IK Solution Rate' is the same feasibility test that defines the reward, so the observed 34% improvement is the policy optimizing its own reward signal. Second, the performance-validation targets are generated via the manipulator's forward kinematics, so they are arm-reachable by definition and do not test whether body posture extends reachability. Some independent evidence does exist: velocity-tracking rewards are roughly maintained, pose error on the generated targets is reported, and zero-shot real-robot demos are shown. These prevent the paper from being wholly circular; however, the load-bearing quantitative claim of 'expanding the workspace by 34%' reduces to the reward oracle's own output. The kinematic-only assumption in Eq. (2) (ignoring leg balance and contact constraints) is a modeling weakness rather than a circularity. No load-bearing self-citation chain was found. Score 6 reflects partial circularity in the headline evaluation, not total equivalence of method and result.

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

The method's central claim rests on hand-tuned reward weights, a numerical IK oracle, and the assumption that manipulator-only kinematics in a body frame captures whole-body reachability. No new physical entities are posited.

free parameters (4)

Reward weight set (Table II) = w_feasible=0.16, w_linear=0.5, w_angular=0.3, w_pose=0.6, w_torque=-1.2e-5, etc.
Hand-tuned coefficients that define the total reward; the paper says regularization weights 'need to be tuned for different legged manipulators.' The orchestration of these weights directly produces the reported behavior.
Feasible-state reward lower bound = 0.2
The term max(e^{sum|ideal_q-q|}, 0.2) in Table II gives a constant positive floor; this floor is chosen by hand and affects exploration pressure.
IK convergence tolerance and iteration count = e_t <= 0.001, max 10 iterations
Algorithm 1 uses these to decide FS=1; changing them changes which torso states are rewarded and also changes the IK solution rate metric.
Command planner target sampling radius = 1.0 m sphere around body frame
Defines the task distribution in training; larger or smaller radii would change difficulty and workspace coverage.

assumptions (4)

domain assumption Reachability of a target depends only on torso pose and manipulator kinematics, not on leg configuration or whole-body constraints.
Used in Eq. (2) to replace full-body feasibility with manipulator-only IK in the body frame. If false, PFG rewards postures the robot cannot actually hold.
domain assumption Damped least-squares IK with 10 iterations and 0.001 threshold is a correct feasibility oracle.
Algorithm 1 computes both the reward and the headline metric; numerical failures count as infeasible.
domain assumption Simulation with domain randomization transfers zero-shot to the real robot.
Hardware results are zero-shot; no real-world adaptation is described.
standard math PPO with the specified observation spaces reliably optimizes this control problem.
Standard RL algorithm; no novel proof or justification is given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Efficient Learning of A Unified Policy For Whole-body Manipulation and Locomotion Skills." pith.science (2026). https://pith.science/paper/ILYIZKUJ

@misc{pith2026250704229,
  author       = {Pith},
  title        = {Pith review of: Efficient Learning of A Unified Policy For Whole-body Manipulation and Locomotion Skills},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ILYIZKUJ}},
  note         = {Machine review of arXiv:2507.04229}
}

read the original abstract

Equipping quadruped robots with manipulators provides unique loco-manipulation capabilities, enabling diverse practical applications. This integration creates a more complex system that has increased difficulties in modeling and control. Reinforcement learning (RL) offers a promising solution to address these challenges by learning optimal control policies through interaction. Nevertheless, RL methods often struggle with local optima when exploring large solution spaces for motion and manipulation tasks. To overcome these limitations, we propose a novel approach that integrates an explicit kinematic model of the manipulator into the RL framework. This integration provides feedback on the mapping of the body postures to the manipulator's workspace, guiding the RL exploration process and effectively mitigating the local optima issue. Our algorithm has been successfully deployed on a DeepRobotics X20 quadruped robot equipped with a Unitree Z1 manipulator, and extensive experimental results demonstrate the superior performance of this approach.

Figures

Figures reproduced from arXiv: 2507.04229 by the authors.

**Figure 1.** A unified learning framework that integrates the kinematic model to effectively guide reinforcement learning, enabling precise whole-body [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

**Figure 2.** Reinforcement learning framework for whole-body locomanipulation. such as velocity tracking [4]–[7] and dexterous manipulation [8]–[11]. Nevertheless, applying DRL to legged manipulators remains challenging due to the interdependent and occasionally conflicting objectives of the arm and body. For instance, to ensure effective manipulation, the quadruped’s body must perform cooperative motions, which can compromis… view at source ↗

**Figure 3.** Physical Feasibility Guided Reward Design. In the first image, the robot is given with a target near the ground and the robot should bend its front legs, facilitating the manipulator’s movement towards the target. The second image depicts the feasible state reward mechanism. When the body explores a feasible state, it is rewarded irrespective of the manipulator’s position. This mechanism will facilitate the learning… view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Command Planner. These two pictures show four smooth reference trajectories generated by the command planner module. Smooth reference trajectories mitigate the learning complexity in the manipulator tracking task while simultaneously enhancing the smoothness of control…

**Figure 5.** Figure 5: Ablation Studies. PFG maintains the robot’s cooperative behavior during locomotion optimization, ensuring the training avoids local optima. trajectory. Fig.4 shows four reference trajectories generated by the command planner. Tref(t) = Tinit ⊕exp 3∆T t 2 total t 2 − 2…

**Figure 7.** Figure 7: Smoothly coordinated whole-body motions in various postures during both static and dynamically configurations. (A-D) illustrates various reaching motions with coordinated whole-body movements demonstrated by the robot during experiments. (E) depicts trajectory trackin…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 18 canonical work pages

[1]

Alma-articulated locomotion and ma- nipulation for a torque-controllable robot,

C. D. Bellicoso, K. Kr ¨amer, M. St ¨auble, D. Sako, F. Jenelten, M. Bjelonic, and M. Hutter, “Alma-articulated locomotion and ma- nipulation for a torque-controllable robot,” in 2019 International conference on robotics and automation (ICRA) . IEEE, 2019, pp. 8477–8483

work page 2019
[2]

A unified mpc framework for whole-body dynamic locomotion and manipula- tion,

J.-P. Sleiman, F. Farshidian, M. V . Minniti, and M. Hutter, “A unified mpc framework for whole-body dynamic locomotion and manipula- tion,” IEEE Robotics and Automation Letters , vol. 6, no. 3, pp. 4688– 4695, 2021

work page 2021
[3]

Versatile multicontact planning and control for legged loco-manipulation,

J.-P. Sleiman, F. Farshidian, and M. Hutter, “Versatile multicontact planning and control for legged loco-manipulation,” Science Robotics, vol. 8, no. 81, p. eadg5014, 2023

work page 2023
[4]

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, p. eaau5872, 2019

2019
[5]

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

2020
[6]

Learning robust perceptive locomotion for quadrupedal robots in the wild,

T. Miki, J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter, “Learning robust perceptive locomotion for quadrupedal robots in the wild,” Science robotics, vol. 7, no. 62, p. eabk2822, 2022

2022
[7]

Dtc: Deep tracking control,

F. Jenelten, J. He, F. Farshidian, and M. Hutter, “Dtc: Deep tracking control,” Science Robotics , vol. 9, no. 86, p. eadh5401, 2024

2024
[8]

Deep reinforcement learn- ing for robotic manipulation with asynchronous off-policy updates,

S. Gu, E. Holly, T. Lillicrap, and S. Levine, “Deep reinforcement learn- ing for robotic manipulation with asynchronous off-policy updates,” in 2017 IEEE international conference on robotics and automation (ICRA). IEEE, 2017, pp. 3389–3396

work page 2017

Show all 26 references

[9]

Scalable deep reinforcement learning for vision-based robotic manipulation,

D. Kalashnikov, et al. , “Scalable deep reinforcement learning for vision-based robotic manipulation,” in Conference on robot learning . PMLR, 2018, pp. 651–673

2018
[10]

A system for general in-hand object re-orientation,

T. Chen, J. Xu, and P. Agrawal, “A system for general in-hand object re-orientation,” in Conference on Robot Learning . PMLR, 2022, pp. 297–307

2022
[11]

Visual dexterity: In-hand reorientation of novel and complex object shapes,

T. Chen, M. Tippur, S. Wu, V . Kumar, E. Adelson, and P. Agrawal, “Visual dexterity: In-hand reorientation of novel and complex object shapes,” Science Robotics , vol. 8, no. 84, p. eadc9244, 2023

2023
[12]

Deep whole-body control: learning a unified policy for manipulation and locomotion,

Z. Fu, X. Cheng, and D. Pathak, “Deep whole-body control: learning a unified policy for manipulation and locomotion,” in Conference on Robot Learning . PMLR, 2023, pp. 138–149

2023
[13]

Learning force control for legged manipulation,

T. Portela, G. B. Margolis, Y . Ji, and P. Agrawal, “Learning force control for legged manipulation,” arXiv preprint arXiv:2405.01402 , 2024

2024 arXiv
[14]

Arm-constrained curriculum learning for loco-manipulation of a wheel-legged robot,

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 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2024, pp. 10 770– 10 776

2024
[15]

Towards a multi-legged mobile manipulator,

B. U. Rehman, M. Focchi, J. Lee, H. Dallali, D. G. Caldwell, and C. Semini, “Towards a multi-legged mobile manipulator,” in 2016 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2016, pp. 3618–3624

2016
[16]

Whole-body control for a torque-controlled legged mobile manipu- lator,

J. Li, H. Gao, Y . Wan, J. Humphreys, C. Peers, H. Yu, and C. Zhou, “Whole-body control for a torque-controlled legged mobile manipu- lator,” in Actuators, vol. 11, no. 11. MDPI, 2022, p. 304

2022
[17]

Teleoperating a legged manipulator through whole-body control,

J. Humphreys, C. Peers, J. Li, Y . Wan, J. Sun, R. Richardson, and C. Zhou, “Teleoperating a legged manipulator through whole-body control,” in Annual Conference Towards Autonomous Robotic Systems . Springer, 2022, pp. 63–77

2022
[18]

Combining learning-based locomotion policy with model-based manipulation for legged mobile manipulators,

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 , vol. 7, no. 2, pp. 2377–2384, 2022

2022
[19]

Proximal policy optimization algorithms,

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

2017 arXiv
[20]

Asymmetric actor critic for image-based robot learning,

L. Pinto, M. Andrychowicz, P. Welinder, W. Zaremba, and P. Abbeel, “Asymmetric actor critic for image-based robot learning,” in 14th Robotics: Science and Systems, RSS 2018 . MIT Press Journals, 2018

2018
[21]

Learning arm-assisted fall damage reduction and recovery for legged mobile manipulators,

Y . Ma, F. Farshidian, and M. Hutter, “Learning arm-assisted fall damage reduction and recovery for legged mobile manipulators,” in 2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 12 149–12 155

2023
[22]

Robotic manipulators and the product of exponentials formula,

R. W. Brockett, “Robotic manipulators and the product of exponentials formula,” in Mathematical Theory of Networks and Systems: Proceed- ings of the MTNS-83 International Symposium Beer Sheva, Israel, June 20–24, 1983 . Springer, 2005, pp. 120–129

1983
[23]

General inverse kinematics with the error damped pseudoinverse,

S. K. Chan and P. D. Lawrence, “General inverse kinematics with the error damped pseudoinverse,” inProceedings. 1988 IEEE international conference on robotics and automation . IEEE, 1988, pp. 834–839

1988
[24]

Numerical method for inverse kine- matics using an extended angle-axis vector to avoid deadlock caused by joint limits,

M. Sekiguchi and N. Takesue, “Numerical method for inverse kine- matics using an extended angle-axis vector to avoid deadlock caused by joint limits,” Advanced Robotics , vol. 35, no. 15, pp. 919–926, 2021

2021
[25]

Manipulator differential kinematics: Part i: Kinematics, velocity, and applications,

J. Haviland and P. Corke, “Manipulator differential kinematics: Part i: Kinematics, velocity, and applications,” IEEE Robotics & Automation Magazine, 2023

2023
[26]

Per-contact iteration method for solving contact dynamics,

J. Hwangbo, J. Lee, and M. Hutter, “Per-contact iteration method for solving contact dynamics,” IEEE Robotics and Automation Letters , vol. 3, no. 2, pp. 895–902, 2018. [Online]. Available: www.raisim.com

2018

Pith tools

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

[1] [1]

Alma-articulated locomotion and ma- nipulation for a torque-controllable robot,

C. D. Bellicoso, K. Kr ¨amer, M. St ¨auble, D. Sako, F. Jenelten, M. Bjelonic, and M. Hutter, “Alma-articulated locomotion and ma- nipulation for a torque-controllable robot,” in 2019 International conference on robotics and automation (ICRA) . IEEE, 2019, pp. 8477–8483

work page 2019

[2] [2]

A unified mpc framework for whole-body dynamic locomotion and manipula- tion,

J.-P. Sleiman, F. Farshidian, M. V . Minniti, and M. Hutter, “A unified mpc framework for whole-body dynamic locomotion and manipula- tion,” IEEE Robotics and Automation Letters , vol. 6, no. 3, pp. 4688– 4695, 2021

work page 2021

[3] [3]

Versatile multicontact planning and control for legged loco-manipulation,

J.-P. Sleiman, F. Farshidian, and M. Hutter, “Versatile multicontact planning and control for legged loco-manipulation,” Science Robotics, vol. 8, no. 81, p. eadg5014, 2023

work page 2023

[4] [4]

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, p. eaau5872, 2019

2019

[5] [5]

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

2020

[6] [6]

Learning robust perceptive locomotion for quadrupedal robots in the wild,

T. Miki, J. Lee, J. Hwangbo, L. Wellhausen, V . Koltun, and M. Hutter, “Learning robust perceptive locomotion for quadrupedal robots in the wild,” Science robotics, vol. 7, no. 62, p. eabk2822, 2022

2022

[7] [7]

Dtc: Deep tracking control,

F. Jenelten, J. He, F. Farshidian, and M. Hutter, “Dtc: Deep tracking control,” Science Robotics , vol. 9, no. 86, p. eadh5401, 2024

2024

[8] [8]

Deep reinforcement learn- ing for robotic manipulation with asynchronous off-policy updates,

S. Gu, E. Holly, T. Lillicrap, and S. Levine, “Deep reinforcement learn- ing for robotic manipulation with asynchronous off-policy updates,” in 2017 IEEE international conference on robotics and automation (ICRA). IEEE, 2017, pp. 3389–3396

work page 2017