Pith. sign in

REVIEW 4 major objections 7 minor 35 references

Reviewed by Pith at T0; open to challenge.

T0 means a machine referee read the full paper against a public rubric. The mark states how deep the mechanical check went, never who wrote it. the ladder, T0–T4 →

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

T0 review · glm-5.2

Fixed-size potential field lets one RL policy throw past any number of obstacles

2026-07-08 07:18 UTC pith:GTBEOOFB

load-bearing objection Potential field grid representation for RL-based robotic throwing in cluttered environments — sound idea, but the empirical case for PFR > EPR is not statistically established the 4 major comments →

arxiv 2607.06388 v1 pith:GTBEOOFB submitted 2026-07-07 cs.RO cs.CVcs.LG

Learning to Throw Objects Safely in Multi-Obstacle Environments

Mohammadreza Kasaei , Klemen Voncina , Hamidreza Kasaei This is my paper
classification cs.RO cs.CVcs.LG
keywords throwingenvironmentsobjectsrepresentationachievesacrossbasketcluttered
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

A robot that throws objects into a basket while avoiding cluttered obstacles needs a state representation that doesn't grow with every new obstacle. This paper introduces the Potential Field Representation (PFR), which encodes the workspace as a fixed 15×15 grid where the basket creates attractive potential and obstacles create repulsive potential. Because the grid size is constant regardless of how many obstacles exist, a single reinforcement learning policy trained on PFR can generalize across arbitrary obstacle counts and configurations—something the alternative Explicit Pose Representation (EPR) cannot do, since EPR appends each obstacle's coordinates to the state vector and thus requires a separate policy per obstacle count. The paper shows that PFR with a CNN encoder and SAC algorithm matches or exceeds EPR's throwing success rates in simulation (above 90% with up to three obstacles), and transfers zero-shot to a real robot achieving 90% success (18/20 throws) with an unseen object (a sneaker) in a cluttered scene. The core mechanism is the combination of a physically meaningful, fixed-dimensional encoding with a CNN that preserves spatial structure, allowing the policy to learn obstacle boundaries and attraction gradients geometrically rather than memorizing obstacle lists.

Core claim

The central object is the Potential Field Representation (PFR): a fixed-size grid encoding where basket attraction and obstacle repulsion are superimposed as continuous potential values, processed by a CNN. The key finding is that this representation simultaneously solves the scalability problem (one policy handles any number of obstacles) and the performance problem (PFR matches or surpasses EPR, which requires obstacle-count-specific policies). SAC emerges as the most consistent RL algorithm for this task. The grid resolution study identifies 15×15 as the best trade-off between spatial fidelity and training cost.

What carries the argument

Potential Field Representation (PFR) on a 15×15 grid with CNN encoder; Soft Actor-Critic (SAC); kinesthetic teaching for safe policy initialization; throwing kernel parameterized by four continuous variables (shoulder joint start/end, release time, motion duration); reward function combining distance-to-goal attraction with obstacle-collision penalty

Load-bearing premise

The real-robot validation uses only 20 trials per condition with a single unseen object and a single obstacle configuration, making the 90% success rate difficult to distinguish from the 70% baseline with statistical confidence.

What would settle it

If PFR-trained policies were tested across substantially more real-robot trials, diverse objects, and obstacle configurations, and the success-rate advantage over EPR disappeared or reversed, the scalability claim would be undermined.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. This paper addresses the problem of learning to throw objects into a target basket while avoiding randomly placed obstacles. The authors propose a Potential Field Representation (PFR) that encodes basket attraction and obstacle repulsion on a fixed-size grid, enabling a single RL policy to handle varying obstacle counts without changing state dimensionality. This is contrasted with an Explicit Pose Representation (EPR) whose dimensionality grows linearly with obstacle count. Policies are initialized via kinesthetic teaching and trained with SAC, DDPG, and TD3 in Gazebo simulation, then transferred zero-shot to a real dual-UR5e robot. The authors evaluate across 4 obstacle counts and 4 objects in simulation, and with 20-trial real-robot experiments using an unseen object (sneaker) under 3 obstacles.

Significance. The core architectural insight—using a fixed-size potential field grid with a CNN encoder to decouple state dimensionality from obstacle count—is sound and practically useful for deployment in unstructured environments. The ablation on grid resolution (Section IV.B) is a valuable design study. The systematic comparison across three RL algorithms and four objects in simulation provides useful empirical breadth. The inclusion of real-robot experiments with an unseen object, while limited in scale, demonstrates sim-to-real feasibility. The work is a reasonable contribution to safe RL for robotic throwing.

major comments (4)
  1. Section IV.A and Table II: PFR is trained for 250k timesteps while EPR is trained for only 100k timesteps. The paper states EPR 'converges' at 100k but provides no learning curves or multi-seed evidence to support this claim. This is load-bearing because the central comparative claim (PFR outperforms EPR) depends on the comparison being fair. If EPR benefits from additional training, the differences in Table II are confounded. The authors should either (a) provide learning curves showing EPR has plateaued at 100k, or (b) train both representations for the same number of timesteps and re-evaluate.
  2. Table II: No variance, confidence intervals, or number of training seeds are reported. Many PFR-vs-EPR differences are small (e.g., SAC/0-obstacles/milk: 0.97 vs 0.95; SAC/1-obstacle/milk: 0.97 vs 0.95). In deep RL, seed-to-seed variance of ±5–10% on success-rate metrics is well documented. Without multi-seed evaluation, it is unclear whether these differences are statistically meaningful. The authors should report results averaged over multiple seeds (≥3) with standard deviations or confidence intervals.
  3. Abstract and Section I claim that PFR enables generalization to 'arbitrary numbers and configurations of obstacles.' However, the evaluation in Table II only covers obstacle counts {0, 1, 3, 5}, all of which are used during training. No experiment tests a PFR policy on an unseen obstacle count (e.g., 2, 4, or 7 obstacles). The scalability claim is architecturally motivated but not empirically validated beyond the training distribution. The authors should either test on unseen obstacle counts or qualify the claim to match what was evaluated.
  4. Table III (real-robot results): The sample size of 20 trials per condition with a single object and single obstacle count is too small to establish the PFR > EPR claim with statistical confidence. The differences (14/20 vs 12/20 for the normal basket; 18/20 vs 14/20 for the large basket) do not reach significance at p<0.05 (Fisher's exact test: p≈0.72 and p≈0.20, respectively). The headline '90% success' figure thus rests on 18/20 trials with a 95% CI of approximately [68%, 99%]. The authors should either increase the sample size or explicitly acknowledge the statistical limitations in the text rather than presenting the differences as confirmed.
minor comments (7)
  1. Section IV.B mentions a 20×20 grid in the quantitative results ('0.97 with a 20×20 grid') but the surrounding text and Figure 5 caption discuss 10×10, 15×15, and 30×30 grids. Please clarify which grid sizes were actually tested.
  2. Section III.F, Eq. (2): The reward function uses two exponential decay terms with exponents -100 and -2. The text says 'e^{-100}' and 'e^{-2}' but the equation writes 'e^{-100d_g^2}' and 'e^{-2d_g^2}'. The notation is clear but the prose description could be more precise about what 'large errors' means in terms of d_g units.
  3. Section III.F, Eq. (3): The penalty term uses d_r defined as the difference between d_g at start and end of episode. The text says 'should this term be positive (the object is closer than it was at the start), d_r = 0.' This clipping logic should be stated more explicitly in the equation or as a separate case.
  4. Table I: The 'epochs' parameter is listed as 50K, but Section IV.A discusses training in terms of timesteps (100k/250k). Please clarify the relationship between epochs and timesteps.
  5. Section IV.D: 'EPR remains was 70%)' contains a grammatical error. Please fix.
  6. The paper references prior work by the authors for the perception pipeline ([27]–[30]) but does not clearly state which components are novel versus reused. A brief clarification would help readers assess the contribution boundary.
  7. Figure 5 caption mentions a 30×30 grid but the quantitative results in the text discuss a 20×20 grid. Please reconcile.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for a careful and constructive review. The referee raises four major concerns: (1) unequal training timesteps for PFR vs. EPR without learning-curve evidence of EPR convergence, (2) absence of multi-seed variance reporting in Table II, (3) an unsubstantiated 'arbitrary obstacle count' generalization claim given that only counts {0,1,3,5} were tested, and (4) insufficient statistical power in the real-robot experiments. We agree that all four points identify genuine weaknesses in the experimental evidence as presented. We will address each through a combination of new experiments and textual revisions.

read point-by-point responses

  1. Referee: Section IV.A and Table II: PFR is trained for 250k timesteps while EPR is trained for only 100k timesteps. The paper states EPR 'converges' at 100k but provides no learning curves or multi-seed evidence to support this claim. This is load-bearing because the central comparative claim (PFR outperforms EPR) depends on the comparison being fair. If EPR benefits from additional training, the differences in Table II are confounded. The authors should either (a) provide learning curves showing EPR has plateaued at 100k, or (b) train both representations for the same number of timesteps and re-evaluate.

    Authors: The referee is correct that the unequal training budget is a confound and that the convergence claim was not substantiated. We will address this in two ways. First, we will include learning curves for both EPR and PFR across all three algorithms, showing reward as a function of training timesteps. These curves will demonstrate that EPR plateaus by approximately 100k steps. Second, as a stronger check, we will train EPR for the full 250k timesteps as well and re-evaluate. If EPR performance remains unchanged or degrades (as can happen with off-policy algorithms due to overfitting or policy collapse), this will confirm that the 100k budget was not disadvantaging EPR. We will report both sets of results in the revised Table II. We note that the architectural argument for PFR—fixed-dimensional state independent of obstacle count—holds regardless of training budget, but we agree the empirical comparison must be fair. revision: yes

  2. Referee: Table II: No variance, confidence intervals, or number of training seeds are reported. Many PFR-vs-EPR differences are small (e.g., SAC/0-obstacles/milk: 0.97 vs 0.95; SAC/1-obstacle/milk: 0.97 vs 0.95). In deep RL, seed-to-seed variance of ±5–10% on success-rate metrics is well documented. Without multi-seed evaluation, it is unclear whether these differences are statistically meaningful. The authors should report results averaged over multiple seeds (≥3) with standard deviations or confidence intervals.

    Authors: We agree. The current Table II reports single-seed results, which is insufficient for the claims made. We will re-run all experiments with at least 5 random seeds per condition and report mean success rates with standard deviations. We will also add a note in the text acknowledging which PFR-vs-EPR differences are and are not statistically significant. For conditions where the differences are within seed variance (such as the SAC/0-obstacle/milk example the referee cites), we will be explicit that PFR and EPR perform comparably, and we will restrict our claims of PFR superiority to conditions where the difference is statistically meaningful. The core contribution of PFR is not that it always outperforms EPR on every cell of the table, but that it achieves comparable or better performance while also being scalable to arbitrary obstacle counts—a property EPR fundamentally lacks. We will sharpen the presentation to reflect this. revision: yes

  3. Referee: Abstract and Section I claim that PFR enables generalization to 'arbitrary numbers and configurations of obstacles.' However, the evaluation in Table II only covers obstacle counts {0, 1, 3, 5}, all of which are used during training. No experiment tests a PFR policy on an unseen obstacle count (e.g., 2, 4, or 7 obstacles). The scalability claim is architecturally motivated but not empirically validated beyond the training distribution. The authors should either test on unseen obstacle counts or qualify the claim to match what was evaluated.

    Authors: The referee is correct that the claim as stated is stronger than what the experiments support. The architectural argument—that PFR's state dimensionality is independent of obstacle count—is valid, but we did not empirically test generalization to unseen obstacle counts. We will take a two-part approach: (1) We will add experiments evaluating the trained PFR policy on unseen obstacle counts of 2, 4, and 7, and report the results in a new table. If performance degrades significantly, we will report this honestly. (2) Regardless of the outcome, we will qualify the language in the abstract and introduction. Specifically, we will replace 'arbitrary numbers and configurations of obstacles' with more precise language such as 'varying numbers and configurations of obstacles within the evaluated range' and note that the architectural design supports scalability in principle, while the empirical validation covers counts 0–7 (or whatever range the new experiments cover). We will not claim generalization beyond what is tested. revision: yes

  4. Referee: Table III (real-robot results): The sample size of 20 trials per condition with a single object and single obstacle count is too small to establish the PFR > EPR claim with statistical confidence. The differences (14/20 vs 12/20 for the normal basket; 18/20 vs 14/20 for the large basket) do not reach significance at p<0.05 (Fisher's exact test: p≈0.72 and p≈0.20, respectively). The headline '90% success' figure thus rests on 18/20 trials with a 95% CI of approximately [68%, 99%]. The authors should either increase the sample size or explicitly acknowledge the statistical limitations in the text rather than presenting the differences as confirmed.

    Authors: The referee's statistical analysis is correct, and we agree that the current sample size is too small to claim a confirmed PFR > EPR advantage on the real robot. We will address this in two ways. First, we will increase the number of real-robot trials to at least 50 per condition (which is feasible given our experimental setup, though time-consuming). This should narrow the confidence intervals substantially. If the differences remain non-significant even with larger samples, we will report that honestly and frame the real-robot results as showing that PFR achieves comparable or numerically higher success rates, without claiming statistical superiority. Second, we will add explicit statistical context to the text: we will report Fisher's exact test p-values and 95% confidence intervals for all real-robot comparisons, and we will add a sentence acknowledging the limitations of the sample size. The headline '90% success' figure will be reported with its confidence interval so readers can assess the uncertainty. We will also soften the language from 'confirm' to 'suggest' where the evidence does not warrant stronger claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity found; the PFR representation is defined independently of the RL training results, and the EPR comparison uses the same RL pipeline.

full rationale

The paper introduces a Potential Field Representation (PFR) defined as a fixed-size grid encoding basket attraction and obstacle repulsion (Section III.C, Fig. 3). This representation is constructed from geometric properties of the environment (positions of basket and obstacles) and is not defined in terms of the RL training outcomes or success rates. The comparison against Explicit Pose Representation (EPR) uses the same RL algorithms (SAC, TD3, DDPG), same reward function (Eq. 1-3), and same training pipeline, differing only in the state encoding. The perception pipeline references prior work by overlapping authors ([27]-[30]), but these are used as external tools for object detection and pose estimation, not as results that the present paper's conclusions depend on. The reward function (Eq. 1-3) is defined in terms of distances in the environment, not in terms of the policy's output or the PFR encoding itself. The claim that PFR generalizes to arbitrary obstacle counts is architecturally true by construction (fixed grid size), and the empirical claim that PFR outperforms EPR is tested through independent simulation and real-robot experiments. While the experimental design has statistical weaknesses (small sample sizes, unequal training timesteps, no seed variation reported), these are correctness concerns rather than circularity: the paper's claims are not equivalent to their inputs by construction. The self-citations to prior perception work are not load-bearing for the central PFR-vs-EPR comparison.

Axiom & Free-Parameter Ledger

11 free parameters · 5 axioms · 0 invented entities

The paper introduces no new physical entities, particles, forces, or mathematical objects beyond standard RL and potential field machinery. The potential field representation is a reformulation of classical artificial potential fields (Khatib 1986) applied as an RL observation. The free parameters are primarily reward shaping constants and standard RL hyperparameters. The axioms are domain assumptions about the task structure and simulation fidelity.

free parameters (11)
  • Reward decay constant (near-goal) = 100
    Exponent in Eq. 2 (e^{-100*d_g^2}), chosen by hand for rapid decay near the goal. No justification provided beyond intuition.
  • Reward decay constant (far-goal) = 2
    Exponent in Eq. 2 (e^{-2*d_g^2}), chosen by hand for gradual improvement. No justification provided.
  • Penalty decay constant = 10
    Exponent in Eq. 3 (e^{-10*d_r^2}), chosen by hand. No justification provided.
  • Reward weight (near-goal) = 0.9
    Weight on the near-goal exponential in Eq. 2. Chosen by hand.
  • Reward weight (far-goal) = 0.1
    Weight on the far-goal exponential in Eq. 2. Chosen by hand.
  • Grid resolution M = 15
    Selected via ablation (Section IV.B) as the best trade-off between training cost and spatial fidelity.
  • Learning rate = 3e-4
    Standard SAC/TD3/DDPG default, not tuned specifically for this task.
  • Discount factor gamma = 0.99
    Standard RL default.
  • Replay buffer size = 50000
    Selected for this task; Table I.
  • Training timesteps (EPR) = 100000
    Chosen based on convergence observation.
  • Training timesteps (PFR) = 250000
    Chosen based on convergence observation for higher-dimensional observation.
axioms (5)
  • domain assumption The throwing task can be modeled as a single-step MDP where one action produces one outcome.
    Section III.E: 'Each training episode consists of a single step, therefore, two observations (before and after) and one action.' This simplifies away sequential decision-making within a throw.
  • domain assumption A 15x15 grid sufficiently captures obstacle geometry for safe throwing trajectories.
    Section IV.B ablation supports this empirically but only for the tested obstacle types and sizes, which are not specified.
  • domain assumption Kinesthetic teaching provides a safe kernel that covers the relevant throwing motion space.
    Section III.B assumes the demonstrated motion is sufficient for RL to modulate; if the kernel is too restrictive, the policy cannot explore certain throwing strategies.
  • domain assumption Gazebo simulation adequately models object dynamics for sim-to-real transfer.
    Section IV.D deploys policies zero-shot, assuming simulation fidelity. The real-robot success rate drop (97% sim to 70-90% real) partially validates this but also shows the gap.
  • domain assumption Obstacles are static during the throwing motion.
    The paper does not consider moving obstacles; the potential field is computed once per episode.

pith-pipeline@v1.1.0-glm · 15232 in / 3333 out tokens · 433797 ms · 2026-07-08T07:18:43.411780+00:00 · methodology

0 comments
read the original abstract

Robotic throwing enables fast and efficient object placement beyond the robot's immediate workspace, but reliable throwing in cluttered environments remains underexplored. Existing approaches, such as TossingBot, learn throwing strategies from visual input but assume obstacle-free settings. In this paper, we address the problem of throwing objects into a target basket while avoiding obstacles placed randomly in the scene. We introduce a potential field state representation that compactly encodes both basket attraction and obstacle repulsion on a fixed-size grid, enabling reinforcement learning (RL) policies to generalize across arbitrary numbers and configurations of obstacles. The policy is initialized from kinesthetic demonstrations and optimized in simulation using three state-of-the-art RL algorithms (SAC, DDPG, TD3). Among these, SAC achieves the most consistent performance across scenarios. We compare the potential field representation against explicit state encodings and demonstrate that it achieves higher success rates and better scalability to unseen obstacle configurations. Real-robot experiments with unseen throwable objects confirm robust sim-to-real transfer, achieving up to $90\%$ success in cluttered scenes. These results demonstrate that PFR provides a practical and robust representation for safe and efficient robotic throwing in unstructured environments. A video showcasing our experiments is available at: https://youtu.be/ZZnJf8ua2dE

Figures

Figures reproduced from arXiv: 2607.06388 by Hamidreza Kasaei, Klemen Voncina, Mohammadreza Kasaei.

Figure 1
Figure 1. Figure 1: Overview of our throwing task. The robot must [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Example of the Potential Field Representation (PFR). [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: Kinesthetic teaching of the throwing kernel. A human demonstrator phys￾ically guides the robot arm through a throwing motion. This provides a safe initial kernel that bootstraps RL training and avoids unsafe random exploration. To avoid unsafe random exploration at the start of RL training, we use kines￾thetic teaching [6]. A hu￾man demonstrator physically guides the robot arm through a throwing motion, wh… view at source ↗
Figure 4
Figure 4. Figure 4: Examples of environment configurations: no obstacles (left), 1 obstacle, 3 obstacles, and 5 obstacles (right). The [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Effect of grid resolution on the Potential Field [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Sequence of snapshots showing the policy successfully delivers the object into the target basket while navigating [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Example output of our real-robot perception system, [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Throwing a sneaker object into the basket while [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages · 3 internal anchors

  1. [1]

    Planning impact-driven logistic tasks,

    A. Zermane, N. Dehio, and A. Kheddar, “Planning impact-driven logistic tasks,”IEEE Robotics and Automation Letters, vol. 9, no. 3, pp. 2184–2191, 2024

    work page 2024

  2. [2]

    Robotic pick-and-toss facilitates urban waste sorting,

    F. Raptopoulos, M. Koskinopoulou, and M. Maniadakis, “Robotic pick-and-toss facilitates urban waste sorting,” in2020 IEEE 16th International Conference on Automation Science and Engineering (CASE). IEEE, 2020, pp. 1149–1154

    work page 2020

  3. [3]

    Toss- ingbot: Learning to throw arbitrary objects with residual physics,

    A. Zeng, S. Song, J. Lee, A. Rodriguez, and T. Funkhouser, “Toss- ingbot: Learning to throw arbitrary objects with residual physics,” in Proceedings of Robotics: Science and Systems (RSS), 2019

    work page 2019

  4. [4]

    Throwing objects into a moving basket while avoiding obstacles,

    H. Kasaei and M. Kasaei, “Throwing objects into a moving basket while avoiding obstacles,” in2023 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 3051–3057

    work page 2023

  5. [5]

    Reinforcement learning to adjust robot movements to new situations,

    J. Kober, E. Oztop, and J. Peters, “Reinforcement learning to adjust robot movements to new situations,” 2011

    work page 2011

  6. [6]

    Trajectories and keyframes for kinesthetic teaching: A human-robot interaction perspective,

    B. Akgun, M. Cakmak, J. W. Yoo, and A. L. Thomaz, “Trajectories and keyframes for kinesthetic teaching: A human-robot interaction perspective,” inProceedings of the seventh annual ACM/IEEE inter- national conference on Human-Robot Interaction, 2012, pp. 391–398

    work page 2012

  7. [7]

    Dual-arm control for coordinated fast grabbing and tossing of an object: Proposing a new approach,

    M. Bombile and A. Billard, “Dual-arm control for coordinated fast grabbing and tossing of an object: Proposing a new approach,”IEEE Robotics & Automation Magazine, vol. 29, no. 3, pp. 127–138, 2022

    work page 2022

  8. [8]

    Event-based Agile Object Catching with a Quadrupedal Robot

    B. Forrai, T. Miki, D. Gehrig, M. Hutter, and D. Scaramuzza, “Event- based agile object catching with a quadrupedal robot,”arXiv preprint arXiv:2303.17479, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  9. [9]

    Bimanual dynamic grabbing and tossing of objects onto a moving target,

    M. Bombile and A. Billard, “Bimanual dynamic grabbing and tossing of objects onto a moving target,”Robotics and Autonomous Systems, p. 104481, 2023

    work page 2023

  10. [10]

    A solution to adaptive mobile manipulator throwing,

    Y . Liu, A. Nayak, and A. Billard, “A solution to adaptive mobile manipulator throwing,” in2022 Ieee/Rsj International Conference On Intelligent Robots And Systems (Iros). IEEE, 2022, pp. 1625–1632

    work page 2022

  11. [11]

    Learning to throw with a handful of samples using decision transformers,

    M. Monastirsky, O. Azulay, and A. Sintov, “Learning to throw with a handful of samples using decision transformers,”IEEE Robotics and Automation Letters, vol. 8, no. 2, pp. 576–583, 2022

    work page 2022

  12. [12]

    Sampling-based motion planning with dynamic intermediate state objectives: Application to throwing,

    Y . Zhang, J. Luo, and K. Hauser, “Sampling-based motion planning with dynamic intermediate state objectives: Application to throwing,” in2012 IEEE International Conference on Robotics and Automation. IEEE, 2012, pp. 2551–2556

    work page 2012

  13. [13]

    Randomized kinodynamic plan- ning,

    S. M. LaValle and J. J. Kuffner Jr, “Randomized kinodynamic plan- ning,”The international journal of robotics research, vol. 20, no. 5, pp. 378–400, 2001

    work page 2001

  14. [14]

    Probabilistically complete kinodynamic planning for robot manipulators with acceleration limits,

    T. Kunz and M. Stilman, “Probabilistically complete kinodynamic planning for robot manipulators with acceleration limits,” in2014 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 2014, pp. 3713–3719

    work page 2014

  15. [15]

    Control for throwing manipulation by one joint robot,

    H. Miyashita, T. Yamawaki, and M. Yashima, “Control for throwing manipulation by one joint robot,” in2009 IEEE International Confer- ence on Robotics and Automation, 2009, pp. 1273–1278

    work page 2009

  16. [16]

    Throwing motion generation using nonlinear optimization on a 6-degree-of-freedom robot manipulator,

    F. Lombai and G. Szederk ´enyi, “Throwing motion generation using nonlinear optimization on a 6-degree-of-freedom robot manipulator,” in2009 IEEE International Conference on Mechatronics. IEEE, 2009, pp. 1–6

    work page 2009

  17. [17]

    Dynamic manipulation,

    M. Mason and K. Lynch, “Dynamic manipulation,” inProceedings of 1993 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS ’93), vol. 1, 1993, pp. 152–159 vol.1

    work page 1993

  18. [18]

    Motion control of a ball throwing robot with a flexible robotic arm,

    Y . Gai, Y . Kobayashi, Y . Hoshino, and T. Emaru, “Motion control of a ball throwing robot with a flexible robotic arm,”International Journal of Computer and Information Engineering, vol. 7, no. 7, pp. 937–945, 2013

    work page 2013

  19. [19]

    Executing optimized throwing motion on robot arm with free joint,

    C. H. Kim and S. Sugano, “Executing optimized throwing motion on robot arm with free joint,”Advanced Robotics, vol. 30, no. 24, pp. 1571–1578, 2016

    work page 2016

  20. [20]

    Curiosity-driven learning of joint locomotion and manipulation tasks,

    C. Schwarke, V . Klemm, M. Van der Boon, M. Bjelonic, and M. Hut- ter, “Curiosity-driven learning of joint locomotion and manipulation tasks,” inProceedings of The 7th Conference on Robot Learning, vol

    work page

  21. [21]

    2594–2610

    PMLR, 2023, pp. 2594–2610

    work page 2023

  22. [22]

    Continuous control with deep reinforcement learning

    T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y . Tassa, D. Silver, and D. Wierstra, “Continuous control with deep reinforce- ment learning,”arXiv preprint arXiv:1509.02971, 2015

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  23. [23]

    Addressing function approxi- mation error in actor-critic methods,

    S. Fujimoto, H. Hoof, and D. Meger, “Addressing function approxi- mation error in actor-critic methods,” inInternational conference on machine learning. PMLR, 2018, pp. 1587–1596

    work page 2018

  24. [24]

    Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,

    T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,” inInternational conference on machine learning. PMLR, 2018, pp. 1861–1870

    work page 2018

  25. [25]

    Ros: an open-source robot operating system,

    M. Quigley, K. Conley, B. Gerkey, J. Faust, T. Foote, J. Leibs, R. Wheeler, A. Y . Ng,et al., “Ros: an open-source robot operating system,” inICRA workshop on open source software, vol. 3, no. 3.2. Kobe, Japan, 2009, p. 5

    work page 2009

  26. [26]

    Robot operating system 2: Design, architecture, and uses in the wild,

    S. Macenski, T. Foote, B. Gerkey, C. Lalancette, and W. Woodall, “Robot operating system 2: Design, architecture, and uses in the wild,” Science Robotics, vol. 7, no. 66, p. eabm6074, 2022

    work page 2022

  27. [27]

    Stable-baselines3: Reliable reinforcement learning implementations,

    A. Raffin, A. Hill, A. Gleave, A. Kanervisto, M. Ernestus, and N. Dormann, “Stable-baselines3: Reliable reinforcement learning implementations,”Journal of Machine Learning Research, vol. 22, no. 268, pp. 1–8, 2021. [Online]. Available: http://jmlr.org/papers/ v22/20-1364.html

    work page 2021

  28. [28]

    Mvgrasp: Real-time multi-view 3d object grasping in highly cluttered environments,

    H. Kasaei and M. Kasaei, “Mvgrasp: Real-time multi-view 3d object grasping in highly cluttered environments,”Robotics and Autonomous Systems, vol. 160, p. 104313, 2023

    work page 2023

  29. [29]

    Simultaneous Multi-View Object Recognition and Grasping in Open-Ended Domains

    H. Kasaei, S. Luo, R. Sasso, and M. Kasaei, “Simultaneous multi- view object recognition and grasping in open-ended domains,”arXiv preprint arXiv:2106.01866, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  30. [30]

    Interactive open-ended object, affordance and grasp learning for robotic manipu- lation,

    S. H. Kasaei, N. Shafii, L. S. Lopes, and A. M. Tom ´e, “Interactive open-ended object, affordance and grasp learning for robotic manipu- lation,” in2019 International Conference on Robotics and Automation (ICRA). IEEE, 2019, pp. 3747–3753

    work page 2019

  31. [31]

    Towards lifelong assistive robotics: A tight coupling between object perception and manipulation,

    S. H. Kasaei, M. Oliveira, G. H. Lim, L. S. Lopes, and A. M. Tom ´e, “Towards lifelong assistive robotics: A tight coupling between object perception and manipulation,”Neurocomputing, vol. 291, pp. 151–166, 2018

    work page 2018

  32. [32]

    Self-supervised 6d object pose estimation for robot manipulation,

    X. Deng, Y . Xiang, A. Mousavian, C. Eppner, T. Bretl, and D. Fox, “Self-supervised 6d object pose estimation for robot manipulation,” in2020 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2020, pp. 3665–3671

    work page 2020

  33. [33]

    Tube acceleration: robust dexterous throwing against release uncertainty,

    Y . Liu and A. Billard, “Tube acceleration: robust dexterous throwing against release uncertainty,”IEEE Transactions on Robotics, 2024

    work page 2024

  34. [34]

    Deep learning approaches to grasp synthesis: A review,

    R. Newbury, M. Gu, L. Chumbley, A. Mousavian, C. Eppner, J. Leit- ner, J. Bohg, A. Morales, T. Asfour, D. Kragic,et al., “Deep learning approaches to grasp synthesis: A review,”IEEE Transactions on Robotics, 2023

    work page 2023

  35. [35]

    Grasp pose detection in point clouds,

    A. Ten Pas, M. Gualtieri, K. Saenko, and R. Platt, “Grasp pose detection in point clouds,”The International Journal of Robotics Research, vol. 36, no. 13-14, pp. 1455–1473, 2017

    work page 2017