Robotics · arXiv:2505.20175 v1 · submitted 26 May 2025 · 38 references
URPlanner: A Universal Paradigm For Collision-Free Robotic Motion Planning Based on Deep Reinforcement Learning
The paper says
“Collision-free motion planning for redundant robot manipulators in complex environments is yet to be explored.”
more of the abstract
Although recent advancements at the intersection of deep reinforcement learning (DRL) and robotics have highlighted its potential to handle versatile robotic tasks, current DRL-based collision-free motion planners for manipulators are highly costly, hindering their deployment and application. This is due to an overreliance on the minimum distance between the manipulator and obstacles, inadequate exploration and decision-making by DRL, and inefficient data acquisition and utilization. In this article, we propose URPlanner, a universal paradigm for collision-free robotic motion planning based on DRL. URPlanner offers several advantages over existing approaches: it is platform-agnostic, cost-effective in both training and deployment, and applicable to arbitrary manipulators without solving inverse kinematics. To achieve this, we first develop a parameterized task space and a universal obstacle avoidance reward that is independent of minimum distance. Second, we introduce an augmented policy exploration and evaluation algorithm that can be applied to various DRL algorithms to enhance their performance. Third, we propose an expert data diffusion strategy for efficient policy learning, which can produce a large-scale trajectory dataset from only a few expert demonstrations. Finally, the superiority of the proposed methods is comprehensively verified through experiments.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 7 Aug 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. Solid engineering paper with a clever new overlap-based reward, but the direct real-world deployment claim is untested and the ED2 demo provenance needs clarification. This line is wrong
3 major objections · 6 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 38 references say.
The paper says, at §I.B and §VII-C
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
The central claim that the trained policy 'can be directly applied to robots in simulators or the real world without requiring fine-tuning' is not supported by the experimental evidence. In §VII-C, the policy is rolled out in the parameterized space in a closed-loop fashion to generate a trajectory, the trajectory is verified for safety, and only then is the joint sequence sent to the real robot for open-loop execution.
more of objection 1
The policy itself is never run on the real robot or in a physics simulator with feedback; the noise-robustness experiments (Fig. 10) are also conducted only in the parameterized environment. Consequently, the fidelity of the abstraction (line-segment manipulator, AABB obstacles, safety offset) to real closed-loop behavior—including unmodeled gripper geometry, tracking errors, and sensor noise—is never validated. This makes the deployment claim an assumption rather than a demonstrated result. I recommend adding closed-loop experiments in a physics simulator (e.g., PyBullet or MuJoCo) or on the real robot with the policy receiving state feedback, and reporting the resulting success rates and collision counts.
The paper says, at §VII-B3, Table VI
The review did not quote this passage.
Pith says: Objection 2 of 3 · major
The comparison between DC-ED2 and DC-RL is not an independent test of ED2's contribution. DC-RL uses expert data collected from a well-trained RL policy on the same task, while DC-ED2 uses trajectories generated by a diffusion model trained on 'a very limited set of collision-free trajectories obtained by an expert agent' (§V-A).
more of objection 2
If these demonstrations are themselves produced by the same (or a closely related) well-trained RL policy, then the observed advantage of DC-ED2 over plain APE2 may simply reflect the use of data from a policy that already solves the task, and the conclusion that ED2 'can significantly enhance DRL training' would be circular. The paper must specify the source of the demonstrations (human teleoperation, scripted policy, or a trained RL agent) and, if they are RL-generated, add a baseline that uses expert data from an independent source (e.g., human demonstrations or a different algorithm) to establish that ED2's diffusion process itself, rather than the origin of the seed data, is what accelerates training.
The paper says, at §VII-C1, safety offset
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
The safety offset ao is selected post hoc: the paper reports that ao=0 cm gives insufficient clearance (MD 1.58 cm), ao=8 cm makes the task infeasible, and ao=5 cm is adopted. Because the feasible range of ao depends on the environment geometry, robot size, and control precision, this is a hand-tuned constant that must be re-selected for each new task, which weakens the claims of universality and cost-effectiveness.
more of objection 3
The paper should either provide a principled procedure for setting ao (for example, derived from the robot's worst-case tracking error or from obstacle uncertainty) or demonstrate that a fixed default value works across a range of environments with different obstacle layouts and manipulator sizes.
6 minor objections
The paper says, at §I.B
Pith says: minor objection 1 of 6
The organization paragraph states that 'some conclusions are given in Section XI', but the manuscript has only nine sections, with conclusions in Section IX. Please correct this cross-reference.
The paper says, at §V-A
Pith says: minor objection 2 of 6
The 'expert agent' that provides the collision-free trajectories for training ED2 is never defined. Please clarify whether these demonstrations come from a human teleoperator, a scripted motion planner, or a pre-trained RL policy, as this is essential for assessing the independence of the ED2 comparison.
show 4 more
The paper says, at §VII-B1, Table II
Pith says: minor objection 3 of 6
The annotation 'Execution Time (s) per 300 steps' is ambiguous. Specify whether this is 300 environment interaction steps or 300 episodes, and clarify the unit of 'steps' in the parameterized environment.
The paper says, at §VII-C1, Fig. 12
Pith says: minor objection 4 of 6
The scoring procedure in Fig. 12 uses min-max normalization, but the normalization ranges are not specified. Please provide the exact formula and the min/max values used for each criterion so that the scores are reproducible.
The paper says, at §VIII-C
Pith says: minor objection 5 of 6
The demonstration strategy shown in Fig. 18 (one demonstration per sub-area, 25 total) appears to be a separate experiment from the main ED2 evaluation, which uses a diffused dataset of 800 trajectories. Please connect this subsection explicitly to the dataset used in Section VII-B3, or state that these are independent evaluations.
The paper says, at §IV-B, Eq. (17)
Pith says: minor objection 6 of 6
The notation in Eq. (17) is unclear: the expression appears to be a dot product between the vector [η, 1−η] and the vector [QLTR, RIR], but the row/column orientation is not defined. Please write the equation in standard linear algebra notation, e.g., VHPE = ηQLTR + (1−η)RIR.
What would settle it
Run the trained APE2:DC-ED2 policy in closed-loop on a simulated Franka Panda in a physics engine (e.g., MuJoCo or PyBullet) with the same tabletop obstacle layout and the same state vector Eq. (2), with no retraining, over the same 50 unseen goal poses from Table V. Record success rate and number of collisions. Also inspect the collision model to confirm whether the gripper/end-effector is included; if it is not, add it and repeat. If the closed-loop success rate falls materially below 100% or any collision occurs, the parameterized-space abstraction does not preserve the real collision-free structure and the direct deployment claim fails.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 7 Aug 2026
- Reader holds up, with conditions
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 6 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
Punchline: URPlanner is a genuine engineering contribution, but the headline claim about direct real-world deployment is not supported by the evidence. Worth reading and worth refereeing, though it needs honest rewording and a couple of experiments or explicit limitations before I'd believe the 'universal' framing.
What's actually new: the UOAR reward based on overlap length between line-segment links and AABB obstacles is a nice replacement for the usual minimum-distance reward. It makes training much cheaper and is genuinely platform-agnostic in the sense that all state and reward computation happens analytically. The APE2 multi-critic scheme with hybrid immediate/long-term evaluation is a reasonable contribution, and the ED2 diffusion-based data augmentation from few demonstrations is clever. The ablations are structured well: they compare rewards, algorithms, and data strategies, and the reported training times (under 9 min vs 10+ hours for simulator-based DRL) are believable given the parameterized environment.
Soft spots, in rough order of importance. First, the stress-test note is right: the 'directly applied to robots... without fine-tuning' claim (Section I.B) is never tested in closed loop. The real-robot experiments in Section VII-C are open-loop executions of pre-planned joint sequences. The policy is rolled out in the parameterized space with feedback, verified, and then replayed on the hardware. Noise robustness (Fig. 10) is also only in the parameterized environment. So the abstraction's fidelity when actually running a reactive policy on a real robot is completely unvalidated. That's the load-bearing gap.
Second, the ED2 evaluation has a mild circularity. The comparison DC-RL vs DC-ED2 pits a diffusion model trained on a few expert demos against a well-trained RL policy's own optimal data. If those demos came from that same RL policy, then you're showing that a distillation of a policy can accelerate training of a similar policy. The paper acknowledges DC-RL is idealized, but it doesn't clarify who provides the demos. If they're human, fine; if they're from a trained RL agent, the 'minimal manual effort' claim weakens. Ask for explicit provenance.
Third, the safety offset ao is tuned post hoc (0, 2.5, 5, 8 cm) and then adopted as 5 cm. They report the trade-off honestly, but it's a per-environment free parameter that could be a source of overfitting in the results.
Minor: no code or data released. For a DRL paper with many hyperparameters, that makes independent verification hard.
Who it's for: researchers in DRL-based manipulation and reward design. It deserves a serious referee—major revision, not desk rejection. The core UOAR idea is worth building on, and the parameterized-space training approach has legs. The claims just need to match the evidence.
The strongest objection
The central claim that the trained policy 'can be directly applied to robots in simulators or the real world without requiring fine-tuning' (Section I.B) is not tested in closed loop. In Section VII-C, URPlanner generates a trajectory by rolling out the policy in the parameterized space, verifies its safety, and then sends the joint sequence to the real robot for open-loop execution. The policy itself is never run on the real robot with feedback; the noise-robustness experiments (Fig. 10) are performed only in the parameterized environment by perturbing the action space. Consequently, the fidelity of the parameterized-space abstraction (line-segment robot, AABB obstacles, expanded safety offset) to the real robot—including unmodeled gripper geometry, control errors, and sensor noise—is never validated under the reactive, closed-loop conditions the paper claims. If the abstraction is even slightly optimistic (e.g., the gripper protrudes beyond the last cylinder, or an AABB under-approximates an obstacle), an open-loop trajectory can collide with no chance for correction. The safety offset ao is also tuned post hoc to 5 cm after observing that 0 cm is unsafe and 8 cm is infeasible, so the method's feasibility depends on a hand-picked constant that must be re-tuned for each new environment. This makes the abstraction the load-bearing assumption, exactly as the reader identified, but the failure mode is specifically the absence of any closed-loop real-world or physics-simulator validation.
The parameterized-space approach is a legitimate engineering advance: replacing minimum-distance rewards with overlap-length rewards is novel, and the analytic environment makes training extremely cheap. The reported numbers (8min49s training, 0.06s trajectory generation, 100% success on the tested tabletop layouts) are internally consistent and credible if the abstraction is exact. However, the paper's headline claim is that the trained policy transfers directly to arbitrary real robots. That claim requires the parameterized space to be a faithful surrogate for the real world under closed-loop operation, including unmodeled geometry, sensor noise, and control errors. The experiments never exercise this: real-robot evaluation is open-loop tracking of a preplanned sequence, and the closed-loop robustness tests are conducted entirely in the parameterized space. A physics-simulator closed-loop test would settle whether the abstraction holds. If it does not, the 'universal paradigm' reduces to a fast offline planner, and the platform-agnostic, direct-deployment advantages would not be established. This concern is consistent with the reader's weakest_assumption, though I emphasize the missing closed-loop validation as the proximate cause. The paper's internal logic is sound; the gap is in the evidence supporting the strongest claim. Therefore, the correct verdict remains conditional on additional validation.
The referee's report
If the parameterized abstraction is faithful, the paper makes a significant practical contribution: it removes the need for physics-simulator interaction during DRL training, reducing training time from hours to minutes and planning time to milliseconds, while providing a dense reward that avoids the computational cost and sensor requirements of minimum-distance computation. The APE2 algorithm, with its multi-critic averaging and early-training reliance on immediate return, is a reasonable and potentially reusable enhancement for deterministic-policy DRL. The ED2 diffusion strategy, if its comparison is made independent, offers an interesting way to amplify a few demonstrations into a large training set. However, the significance is currently conditional on two load-bearing assumptions that are not adequately tested: that the line-segment/AABB parameterization preserves the collision-free structure of real manipulation tasks in closed-loop execution, and that the expert demonstrations used by ED2 are not themselves the source of the observed acceleration.
Pith's reading of the paper (a machine's summary, not the authors')
URPlanner aims to make deep-reinforcement-learning motion planning for robot manipulators cheap enough to use in practice: training in minutes, planning in milliseconds, with no inverse kinematics and no simulated or real platform in the loop. The paper's central bet is that a parameterized task space — robot links as line segments inside cylinders, obstacles as expanded axis-aligned boxes — lets a DRL agent learn collision-free reaching in a purely analytic environment, then transfer directly to real robots. Against that background, it introduces three components: a dense overlap-based reward that avoids minimum-distance calculations, an exploration/evaluation wrapper (APE2) that improves DDPG/TD3/E3AC, and a diffusion-based strategy (ED2) that manufactures a large expert-trajectory dataset from a handful of demonstrations. The reported experiments support the claim with 100% testing success on tabletop tasks, sub-9-minute training, 0.06-second trajectory generation, and better trajectory quality than RRT/RRT* and prior DRL planners. The paper argues this combination removes the main cost barrier to deploying learned motion planners on arbitrary manipulators.
What carries the argument. The load-bearing object is the parameterized task space: a manipulator link $i$ is the line segment $s_{i-1}^{i} = p_{i-1} + \overrightarrow{p_{i-1}p_i}\,\lambda$, $0 \le \lambda \le 1$, enclosed by a cylinder of radius $a_r$, and each obstacle is an axis-aligned bounding box expanded by $a_r$ plus a safety offset $a_o$. Collision avoidance is then the statement that no segment intersects any expanded box, and the dense reward is $r_{\mathrm{UOAR}} = -\frac{\sum_j\sum_g l_{jg}}{\sum_j L_j}$, the total overlap length between segments and boxes normalized by total segment length. This turns minimum-distance computation into an intersection-length calculation via the slab method (Eqs. 7-8), which is what makes training analytically computable and platform-agnostic. The APE2 wrapper (action candidates from $M \times N$ Gaussian noises; hybrid evaluation $\eta\,Q_{\mathrm{LTR}} + (1-\eta)R_{\mathrm{IR}}$ with $\eta = \mathrm{clip}(t_c/T, 0, 1)$) and ED2 diffusion over trajectories are the supporting mechanisms that accelerate and stabilize learning on top of that geometry.
Core claim. The paper claims that collision-free motion planning for arbitrary redundant manipulators reduces to a geometric overlap problem that DRL can solve cheaply. Instead of computing the minimum distance between robot and obstacles, URPlanner expands each obstacle's axis-aligned bounding box by the robot's link-cylinder radius plus a safety offset, represents each link as a line segment, and uses the normalized overlap length between segments and boxes as a dense reward. Because forward kinematics and reward are evaluated analytically from DH parameters in this parameterized space, no simulator, real robot, or IK solver is needed for training; the same parameterized rollout produces the executed trajectory. The paper argues this foundation, combined with APE2's multi-action exploration and hybrid Q/immediate-return evaluation and ED2's diffusion-generated expert data, yields a planner that is platform-agnostic, trains in minutes, plans in tens of milliseconds, and transfers without fine-tuning.
The premise it rests on. The whole approach rests on the proxy geometry being faithful: if a real link pokes outside its modeled cylinder, or a real obstacle bulges outside its expanded box, a trajectory that is collision-free in the proxy world could still hit something on the actual robot.
If the paper is right (Pith's inferences, not the authors')
A new manipulator can be planned for with no simulator build, no inverse-kinematics solver, and no platform interaction: DH parameters and link radii define the training environment.
Because state transitions and rewards are analytic, a full trajectory is emitted in about 60 ms, which makes closed-loop replanning at a few hertz practical for moving goals.
The overlap reward keeps per-interaction cost nearly flat as obstacles increase (0.78 to 0.86 s per 300 interactions from 4 to 16 obstacles), whereas min-distance PLR scales from 9.4 to 32.8 s.
Non-optimal, diffusion-generated expert data accelerate early DRL training without capping final performance, so a few demonstrations per region suffice for fast policy learning.
Trained sub-policies can be stored in a library and adapted to new layouts via transfer learning in roughly one to three minutes.
A testable extension the paper leaves implicit would condition the ED2 diffusion model on goal-pose or obstacle-layout embeddings, letting one model generate expert data for the whole workspace instead of per-region demonstration sets.
The same overlap reward could be carried into dynamic scenes by updating bounding boxes each time step; the paper flags dynamic environments as future work but does not test this, and the analytic environment makes the benchmark cheap to run.
The 0.06-second figure describes open-loop trajectory generation in the proxy space; a closed-loop comparison that includes real robot state feedback would give a sharper deployment-cost number than the paper reports.
An adaptive safety offset might extend the method into narrow passages, since the paper shows a fixed 8 cm offset makes its test task infeasible while 0 cm sacrifices clearance.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Collision-free motion planning for redundant robot manipulators in complex environments is yet to be explored. Although recent advancements at the intersection of deep reinforcement learning (DRL) and robotics have highlighted its potential to handle versatile robotic tasks, current DRL-based collision-free motion planners for manipulators are highly costly, hindering their deployment and application. This is due to an overreliance on the minimum distance between the manipulator and obstacles, inadequate exploration and decision-making by DRL, and inefficient data acquisition and utilization. In this article, we propose URPlanner, a universal paradigm for collision-free robotic motion planning based on DRL. URPlanner offers several advantages over existing approaches: it is platform-agnostic, cost-effective in both training and deployment, and applicable to arbitrary manipulators without solving inverse kinematics. To achieve this, we first develop a parameterized task space and a universal obstacle avoidance reward that is independent of minimum distance. Second, we introduce an augmented policy exploration and evaluation algorithm that can be applied to various DRL algorithms to enhance their performance. Third, we propose an expert data diffusion strategy for efficient policy learning, which can produce a large-scale trajectory dataset from only a few expert demonstrations. Finally, the superiority of the proposed methods is comprehensively verified through experiments.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-09-23
The authors' main claim, as the classifier picked it out 0.93
“Finally, the superiority of the proposed methods is comprehensively verified through experiments.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.36 · states limitation 0.09 · human participants 0.03 · controlled comparison 0.23 · randomized assignment 0.01 · states quantitative result 0.04 · compares with prior results 0.64 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Collision-free motion planning for redundant robot manipulators in complex environments is yet to be explored. | gap 0.93 | 0.05 | 0.03 | asserted 0.97 |
| Although recent advancements at the intersection of deep reinforcement learning (DRL) and robotics have highlighted its potential to handle versatile robotic tasks, current DRL-based collision-free motion planners for manipulators are highly costly, hindering their deployment and application. | background 0.77 | 0.05 | 0.04 | asserted 0.96 |
| This is due to an overreliance on the minimum distance between the manipulator and obstacles, inadequate exploration and decision-making by DRL, and inefficient data acquisition and utilization. | background 0.66 | 0.08 | 0.03 | asserted 1.00 |
| In this article, we propose URPlanner, a universal paradigm for collision-free robotic motion planning based on DRL. | contribution 1.00 | 0.83 | 0.19 | asserted 0.99 |
| URPlanner offers several advantages over existing approaches: it is platform-agnostic, cost-effective in both training and deployment, and applicable to arbitrary manipulators without solving inverse kinematics. | contribution 0.99 | 0.63 | 0.23 | asserted 1.00 |
| To achieve this, we first develop a parameterized task space and a universal obstacle avoidance reward that is independent of minimum distance. | method 0.92 | 0.92 | 0.05 | asserted 0.99 |
| Second, we introduce an augmented policy exploration and evaluation algorithm that can be applied to various DRL algorithms to enhance their performance. | method 0.61 | 0.92 | 0.04 | asserted 0.99 |
| Third, we propose an expert data diffusion strategy for efficient policy learning, which can produce a large-scale trajectory dataset from only a few expert demonstrations. | method 0.69 | 0.93 | 0.05 | asserted 0.97 |
| Finally, the superiority of the proposed methods is comprehensively verified through experiments. | result 0.81 | 0.90 | 0.56 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.82
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 6,907 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (19)
Figures are the authors', reproduced from the paper.
References (38)
- [1] J. Lv, Y . Wang, S. Wang, X. Bai, R. Wang, and M. Tan, “A collision-free planning and control framework for a biomimetic underwater vehicle in dynamic environments,” IEEE/ASME Transactions on Mechatronics , vol. 28, no. 3, pp. 1415–1424, 2023 on Pith
- [2] M. Safeea, R. Bearee, and P. Neto, “A modified DLS scheme with controlled cyclic solution for inverse kinematics in redundant robots,” IEEE Transactions on Industrial Informatics , vol. 17, no. 12, pp. 8014– 8023, Dec., 2021 on Pith
- [3] A. H. Qureshi, Y . Miao, A. Simeonov, and M. C. Yip, “Motion planning networks: Bridging the gap between learning-based and classical motion planners,” IEEE Transactions on Robotics , vol. 37, no. 1, pp. 48–66, 2021 on Pith
- [4] S. Ruan, K. L. Poblete, H. Wu, Q. Ma, and G. S. Chirikjian, “Efficient path planning in narrow passages for robots with ellipsoidal compo- nents,” IEEE Transactions on Robotics , vol. 39, no. 1, pp. 110–127, 2023 on Pith
- [5] X. Li, H. Liu, and M. Dong, “A general framework of motion planning for redundant robot manipulator based on deep reinforcement learning,” IEEE Transactions on Industrial Informatics , vol. 18, no. 8, pp. 5253– 5263, 2022 on Pith
- [6] A. Faust, K. Oslund, O. Ramirez, A. Francis, L. Tapia, M. Fiser, and J. Davidson, “PRM-RL: Long-range robotic navigation tasks by combin- ing reinforcement learning and sampling-based planning,” in 2018 IEEE International Conference on Robotics and Automation (ICRA), 2018, pp.... on Pith
- [7] H. Liu, F. Ying, R. Jiang, Y . Shan, and B. Shen, “Obstacle-avoidable robotic motion planning framework based on deep reinforcement learn- ing,” IEEE/ASME Transactions on Mechatronics , vol. 29, no. 6, pp. 4377–4388, 2024 on Pith
- [8] F. Ying, H. Liu, R. Jiang, and M. Dong, “Extensively explored and evaluated actor-critic with expert-guided policy learning and fuzzy feedback reward for robotic trajectory generation,” IEEE Transactions on Industrial Informatics , vol. 18, no. 11, pp. 7749–7760, 2022 on Pith
- [9] F. Ying, H. Liu, R. Jiang, and X. Yin, “Trajectory generation for multi- process robotic tasks based on nested dual-memory deep deterministic policy gradient,” IEEE/ASME Transactions on Mechatronics , vol. 27, no. 6, pp. 4643–4653, 2022 on Pith
- [10] S. Safaoui, A. P. Vinod, A. Chakrabarty, R. Quirynen, N. Yoshikawa, and S. D. Cairano, “Safe multiagent motion planning under uncertainty for drones using filtered reinforcement learning,” IEEE Transactions on Robotics, vol. 40, pp. 2529–2542, 2024 on Pith
- [11] R. Park, J. Kdela, R. Matouek, and M. Juek, “Deep-reinforcement- learning-based motion planning for a wide range of robotic structures,” Computation, vol. 12, no. 6, 2024. [Online]. Available: https://www.mdpi.com/2079-3197/12/6/116 on Pith
- [12] Z. Hu, Y . Zheng, and J. Pan, “Grasping living objects with adversarial behaviors using inverse reinforcement learning,” IEEE Transactions on Robotics, vol. 39, no. 2, pp. 1151–1163, 2023 on Pith
- [13] A. Zeng, S. Song, S. Welker, J. Lee, A. Rodriguez, and T. Funkhouser, “Learning synergies between pushing and grasping with self-supervised deep reinforcement learning,” in 2018 IEEE/RSJ International Confer- ence on Intelligent Robots and Systems (IROS) , 2018, pp. 4238–4245 on Pith
- [14] Y . Deng, X. Guo, Y . Wei, K. Lu, B. Fang, D. Guo, H. Liu, and F. Sun, “Deep reinforcement learning for robotic pushing and picking in cluttered environment,” in 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2019, pp. 619–626 on Pith
- [15] L. Berscheid, P. Meiner, and T. Krger, “Robot learning of shifting objects for grasping in cluttered environments,” in 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2019, pp. 612– 618 on Pith
- [16] Z. Hou, Z. Li, C. Hsu, K. Zhang, and J. Xu, “Fuzzy logic-driven variable time-scale prediction-based reinforcement learning for robotic multiple peg-in-hole assembly,” IEEE Transactions on Automation Science and Engineering, vol. 19, no. 1, pp. 218–229, 2022 on Pith
- [17] Z. Hou, W. Yang, R. Chen, P. Feng, and J. Xu, “A hierarchical compliance-based contextual policy search for robotic manipulation tasks with multiple objectives,” IEEE Transactions on Industrial Infor- matics, vol. 19, no. 4, pp. 5444–5455, 2023. 20 on Pith
- [18] H. Xu, Y . Luo, S. Wang, T. Darrell, and R. Calandra, “Towards learning to play piano with dexterous hands and touch,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2022, pp. 10 410–10 416 on Pith
- [19] D. Bianchi, M. G. Antonelli, C. Laschi, A. M. Sabatini, and E. Falotico, “Softoss: Learning to throw objects with a soft robot,” IEEE Robotics and Automation Magazine , pp. 2–12, 2023 on Pith
- [20] L. Pecyna, S. Dong, and S. Luo, “Visual-tactile multimodality for following deformable linear objects using reinforcement learning,” in 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022, pp. 3987–3994 on Pith
show 18 more
- [21] Y . Chen, F. Ying, X. Li, and H. Liu, “Deep reinforcement learning in maximum entropy framework with automatic adjustment of mixed temperature parameters for path planning,” in 2023 7th International Conference on Robotics, Control and Automation (ICRCA) , 2023, pp. 78–82 on Pith
- [22] Z. Sui et al. , “Formation control with collision avoidance through deep reinforcement learning using model-guided demonstration,” IEEE Transactions on Neural Networks and Learning Systems , vol. 32, no. 6, pp. 2358–2372, 2021 on Pith
- [23] W. Liu, L. Peng, J. Cao, X. Fu, Y . Liu, Z. Pan, and J. Yang, “Ensemble bootstrapped deep deterministic policy gradient for vision-based robotic grasping,” IEEE Access, vol. 9, pp. 19 916–19 925, 2021 on Pith
- [24] J. Wu, R. Wang, R. Li, H. Zhang, and X. Hu, “Multi-critic DDPG method and double experience replay,” in 2018 IEEE International Conference on Systems, Man, and Cybernetics (SMC) , 2018, pp. 165– 171 on Pith
- [25] Y . Ma, D. Xu, and F. Qin, “Efficient insertion control for precision assembly based on demonstration learning and reinforcement learning,” IEEE Transactions on Industrial Informatics , vol. 17, no. 7, pp. 4492– 4502, 2021 on Pith
- [26] P. Chen, J. Pei, W. Lu, and M. Li, “A deep reinforcement learning based method for real-time path planning and dynamic obstacle avoidance,” Neurocomputing, vol. 497, pp. 64–75, 2022 on Pith
- [27] A. Xie, Y . Lee, P. Abbeel, and S. James, “Language-conditioned path planning,” 2023. [Online]. Available: https://arxiv.org/abs/2308.16893 on PithPith's reviewarXiv
- [28] Z. Hou, J. Fei, Y . Deng, and J. Xu, “Data-efficient hierarchical rein- forcement learning for robotic assembly control applications,” IEEE Transactions on Industrial Electronics , vol. 68, no. 11, pp. 11 565– 11 575, 2021 on Pith
- [29] Z. Mu, W. Xu, X. Gao, L. Xue, and C. Li, “Obstacles modeling and collision detection of space robots for performing on-orbit services,” in 2014 4th IEEE International Conference on Information Science and Technology, 2014, pp. 461–466 on Pith
- [30] S. Yang and Q. Wang, “Robotic arm motion planning with autonomous obstacle avoidance based on deep reinforcement learning,” in 2022 41st Chinese Control Conference (CCC) , 2022, pp. 3692–3697 on Pith
- [31] Y .-L. Kim, K.-H. Ahn, and J.-B. Song, “Reinforcement learning based on movement primitives for contact tasks,” Robotics and Computer- Integrated Manufacturing, vol. 62, p. 101863, 2020 on Pith
- [32] J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” in Advances in Neural Information Processing Systems, vol. 33. Curran Associates, Inc., 2020, pp. 6840– on Pith
- [33] M. Janner, Y . Du, J. Tenenbaum, and S. Levine, “Planning with diffusion for flexible behavior synthesis,” in International Conference on Machine Learning, 2022 on Pith
- [34] H. Liu, J. Tan, Y . Chen, and M. H. Ang, “Learning complicated ma- nipulation skills via deterministic policy with limited demonstrations,” in 2023 21st International Conference on Advanced Robotics (ICAR) , 2023, pp. 499–505 on Pith
- [35] P. J. Ball, L. Smith, I. Kostrikov, and S. Levine, “Efficient online reinforcement learning with offline data,” in Proceedings of the 40th International Conference on Machine Learning , 2023 on Pith
- [36] J. Xu, Z. Hou, W. Wang, B. Xu, K. Zhang, and K. Chen, “Feedback deep deterministic policy gradient with fuzzy reward for robotic multiple peg- in-hole assembly tasks,” IEEE Transactions on Industrial Informatics , vol. 15, no. 3, pp. 1658–1667, 2019 on Pith
- [37] Z. Zhang, J. Hong, A. M. S. Enayati, and H. Najjaran, “Using implicit behavior cloning and dynamic movement primitive to facilitate rein- forcement learning for robot motion planning,” IEEE Transactions on Robotics, vol. 40, pp. 4733–4749, 2024. Fengkang Ying (Graduate Student... on Pith
- [6851] [Online]. Available: https://proceedings.neurips.cc/paper/2020/ file/4c5bcfec8584af0d967f1ab10179ca4b-Paper.pdf on Pith
Cited by 2 Pith papers
Reviewed papers on Pith that cite this one. The full citation record
How the literature uses it
What this paper does with the arXiv papers it cites
What it cites
Cites as background 1 paper
-
For manipulators, a motion planning problem is to find a sequence of collision-free robot configurations represented as waypoints and the corresponding joint angles [ref] .
Language-Conditioned Path Planning · Problem Formulation · 0.82
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment