{"id":"eca83576-a7b6-4706-959d-a8c8c0dd6fd2","arxiv_id":"1908.03440","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A simulated DRL agent learns to position a suction gripper over blocks using only depth-camera images, but the claimed precision is not numerically demonstrated.","lead":"The paper trains deep reinforcement learning agents to grasp blocks in a simulated Unity environment using depth camera images as the only sensor input. It reports that PPO with curriculum learning learns to position a vacuum gripper near targets, but provides no quantitative grasp success or error measurements.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central 'negligible errors' claim rests on a reward-proxy that is never checked against physical grasp success or direct pose-error measurements.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the hand-defined reward thresholds are used as a proxy for grasp success, but no direct measurement of tool-pose error or physical grasping outcome is provided. I agree with that assessment. The internal consistency of the learning curves is not the issue; the problem is the external validity of the reward proxy. A policy can maximize the dense reward while still failing to achieve suction contact or lift, and the paper provides no evidence to rule out that failure mode. The paper also lacks error bars, a fixed evaluation protocol, and released code or data, all of which limit the support for the abstract's strong claim. There is no machine-checked proof or independent reproduction to offset these gaps. Therefore, the central claim is unsubstantiated by the reported experiments, and the reader's REJECT verdict should stand unchanged.","tokens_in":9206,"tokens_out":2869,"duration_ms":32256,"concrete_test":"Re-run the reported PPO+curriculum configuration in the Unity environment on 500 held-out random block configurations. For each episode, log the final tooltip position and y-rotation error against the block surface, plus a physical-suction flag indicating whether the gripper maintains contact and lifts the block. Report median and 95th-percentile errors and the grasp success rate. If reward-flagged successes do not coincide with actual lift successes, or if the final error distribution is not tightly within the final lesson thresholds, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is that the dense reward thresholds defined in Section III-C are a faithful proxy for a successful suction grasp. A positive reward is emitted when the tooltip is within ±0.1 m in x/y, [0.01, 0.02] m in z, and ±10° in y-rotation, with the curriculum in Section III-D finally tightening this to ±0.01 m and ±2°. The abstract's claim of 'negligible position and orientation errors' and of a 'successful attempt' is supported only by cumulative-reward curves in Figure 5. No held-out evaluation reports actual tooltip errors, contact events, suction engagement, or whether the object was lifted. Since the reward definition is exactly the quantity claimed as solved, a high cumulative reward only demonstrates that the policy satisfied the proxy; it does not demonstrate that the proxy corresponds to physically successful grasps. For a vacuum gripper, success additionally requires correct surface contact and normal alignment; the z window [0.01, 0.02] m is not tied to any measured suction or lift outcome. The experiment-termination policy in Section V ('tests were interrupted when the result seemed satisfactory') further weakens any quantitative inference. The core 'negligible errors' claim is therefore unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a deep reinforcement learning (DRL) approach for controlling a simulated Kuka KR16 robot equipped with a vacuum gripper, using 2.5D depth images as the only visual input. The goal is to regress a 3D end-effector pose that brings the suction head to the surface of blocks of varying shape, size, position, and orientation. The authors compare PPO, TRPO, and DDPG, and also study curriculum learning, different CNN architectures, and input configurations (depth only, depth plus RGB, frame grab). The reported results are cumulative reward curves during training. The abstract claims that the method reaches the object surface with \"negligible position and orientation errors\" and that this is \"the first successful attempt\" of using 2.5D images only as input to a DRL algorithm for grasping with 3D world-coordinate regression.","tokens_in":9407,"tokens_out":3010,"duration_ms":32168,"significance":"If the central claim were supported, the work would be a useful demonstration that a DRL policy can regress a precise suction-grasp pose from depth images alone, potentially enabling low-cost visual grasping in structured environments. The paper also provides a simulated environment and a comparison of popular DRL algorithms on this task, which could be a starting point for further study. However, the significance as currently established is low: the only evidence presented consists of training reward curves, with no quantitative measurement of pose error, grasp success, or physical validity of the learned behavior. The claimed novelty and effectiveness are therefore not demonstrated by the experiments.","major_comments":[{"comment":"The central claim of \"negligible position and orientation errors\" is not supported by any quantitative measurement. All results shown in Figure 5 are cumulative reward curves during training, and no held-out evaluation reports tooltip error, percentage of successful grasps, contact events, or whether the object was lifted. A rising training reward is an optimization signal, not a performance metric, so the abstract's claim is unsubstantiated by the experimental section.","section":"Abstract and Section V"},{"comment":"The reward function defines success as reaching hand-set thresholds: ±0.1 m in x/y, 0.01–0.02 m in z, and ±10 degrees in y-rotation, tightened in the curriculum to ±0.01 m and 2 degrees. The paper never validates that these thresholds correspond to a physically successful suction grasp. For a vacuum gripper, success requires proper surface contact and normal alignment, and the z window in particular is not tied to any measured suction or lift outcome. Thus a high cumulative reward only shows that the policy satisfied the proxy, not that the proxy is faithful.","section":"Section III-C and Section III-D"},{"comment":"The experimental protocol is not rigorous enough to support the paper's conclusions. The authors state that \"tests were interrupted when the result seemed satisfactory,\" and no seed variance, multiple runs, or statistical analysis is reported. With a single run and subjective stopping criteria, the comparison between algorithms (PPO good, DDPG poor, TRPO unstable) is anecdotal rather than quantitative. The paper should report held-out metrics with confidence intervals or at least multiple seeds.","section":"Section V"},{"comment":"The claim of being \"the first successful attempt\" is unsupported because no comparison to existing grasp methods or baselines is provided, and there is no quantitative evidence that the proposed method succeeds at grasping. The novelty claim cannot be evaluated from the reported experiments.","section":"Abstract and Section I"}],"minor_comments":[{"comment":"Equation (1) uses 'actualv' without defining how this value is obtained from the raw sensor reading; clarify the notation and units.","section":"Section III-A, Eq. (1)"},{"comment":"The reference \"stable baselines [?]\" is missing its citation; please add the appropriate reference.","section":"Section V"},{"comment":"There are numerous typos and grammatical errors, e.g., \"as of the input\" in the abstract, \"particularity adapt\" in Section V-C, and \"starting form\" in Section III-D. A careful proofreading is needed.","section":"Throughout"},{"comment":"The caption of Figure 5 lists subfigures (a)–(k) but does not describe the content of each subfigure; the reader must infer which configuration each curve corresponds to from the text, which is difficult.","section":"Figure 5"},{"comment":"The total reward function is not explicitly defined as a sum of the sparse terms (0.1, -0.1, 0.5) and the dense terms r_fmt and r_fft; the combination rule should be stated explicitly.","section":"Section III-C"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an early preprint. The main issue is not a minor flaw but a complete absence of quantitative evaluation of the claimed grasp performance. The central claims—negligible errors and first successful 2.5D-only DRL grasping—are unsupported by the presented evidence. I recommend rejection, though the authors could potentially resubmit after adding substantial new experiments with held-out metrics, baseline comparisons, and statistical analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this paper does one real thing and then undercuts it with weak evaluation. It applies PPO, TRPO, and DDPG to regress a 3D suction-pick pose from depth images only, in a Unity simulation with a Kuka arm. That combination—depth-only input, continuous 3D pose output, vacuum gripper—is not something I've seen in the cited DRL-grasping literature, so as an incremental proof-of-concept it has a legitimate place.\n\nThe reward design is thoughtful: dense rewards based on tooltip position and y-rotation, plus curriculum learning that tightens the target region over 19 lessons. The authors also test several CNN input sizes, frame grab, and an RGB+depth variant. That is useful engineering detail.\n\nThe soft spot is the one that matters. The paper's abstract claims the policy reaches the object surface with 'negligible position and orientation errors' and calls this 'the first successful attempt.' The experiments never measure pose error on a held-out set, never report contact events, suction engagement, or whether a block is lifted. The only evidence is cumulative reward curves. Because the reward is exactly defined by position/orientation thresholds, a rising reward curve shows the policy satisfied the proxy, not that the proxy corresponds to a physical grasp. For a vacuum gripper, contact alignment matters; a 1 cm z-window and 2-degree rotation threshold are not tied to any measured lift. In fact, the paper admits tests were stopped when results 'seemed satisfactory,' which further weakens inference. This is a textbook case of evaluating the reward proxy as if it were the task.\n\nAlso missing: baselines from classical vision or depth-reading methods, variance over seeds, and released code or data. The related-work section is thin and doesn't compare against depth-image grasping methods that don't use DRL. The 'first successful attempt' claim is too strong given no quantitative evaluation.\n\nProportionate verdict: as a proof-of-concept, this is plausible and worth knowing about, but the central claim is unverified. It might be a good workshop paper if rewritten with honest claims and followed by real experiments. As it stands, I would not send it to a serious peer-review venue; the evidence is too thin. A desk reject with an invitation to resubmit with direct pose-error and grasp-success measurements would be the right call.\n\nThe paper is for people specifically working on DRL-based grasping in simulation, as a cautionary example about reward proxies. I wouldn't cite it, but I might use it in a reading group discussion about evaluation practices.","headline":"A plausible depth-only DRL proof-of-concept whose headline 'negligible errors' claim is not backed by the reported experiments.","tokens_in":9959,"tokens_out":2921,"would_cite":false,"duration_ms":31039,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A deep reinforcement learning policy trained on 2.5D depth images alone can regress a suction-grasp tool pose with negligible position and orientation errors, according to this paper.","keywords":["deep reinforcement learning","grasping","2.5D images","depth camera","vacuum gripper","curriculum learning","PPO","robot manipulation"],"falsifier":"Evaluate the trained PPO-plus-curriculum policy on a held-out set of random block positions, orientations, and shapes, recording the physical outcome of the suction attempt (whether the block lifts and stays attached) together with the measured tooltip pose error. If the policy frequently enters the rewarded tolerance zone without achieving a grasp, the dense-reward proxy is the weak link; if it misses the zone, the pose-regression claim fails.","tokens_in":8985,"feed_emoji":"🤖","tokens_out":6589,"duration_ms":64200,"temperature":0.7,"pith_summary":"This paper claims that a deep reinforcement learning agent can grasp flat blocks with a vacuum gripper using only 2.5D depth images as input, and that this is the first demonstration of a DRL policy regressing 3D tool coordinates from such images alone. The authors built a simulated robot setup with a fixed depth camera and compared PPO, TRPO, and DDPG. They report that PPO combined with a 19-lesson curriculum lands the suction tooltip on the block surface with position error near 1 cm and yaw error near 2 degrees, while the other algorithms underperform. If true, the result matters because depth-only perception is cheap and robust to illumination changes, making flexible suction grasping easier to deploy in structured settings.","feed_headline":"Robot learns suction grasps from depth images alone","feed_subtitle":"In simulation, PPO plus a 19-lesson curriculum beats TRPO and DDPG, landing within centimeters and 2 degrees.","key_machinery":"The carrying mechanism is the pairing of dense reward shaping with a 19-lesson curriculum on a PPO policy. The state is a single simulated depth image, with intensity scaled to a 0.4-2 m camera range and Gaussian noise added; the action is a continuous tool pose. Rewards grant +0.1 for touching the target, -0.1 for undesired collisions, +0.5 for reaching the position tolerance around the block centroid, and another +0.5 for matching the block's yaw, plus two progressive rewards that rate the direction of motion and end-effector orientation. The curriculum narrows the position tolerance from 0.1 m toward 0.01 m and the yaw tolerance from 10 degrees to 2 degrees across lessons, pushing the policy toward precision.","core_discovery":"On the paper's own terms, the central discovery is that a Proximal Policy Optimization (PPO) agent can solve a continuous suction-grasping task when fed a single 2.5D depth image and asked to output an end-effector pose. The final curriculum lesson demands the tooltip be within 0.01 m of the block centroid in each positional coordinate and within 2 degrees of the block's yaw rotation, and the paper reports that the policy reaches the object surface with negligible position and orientation errors. The same experiments show DDPG performing poorly, TRPO being unstable, and PPO without curriculum learning not reaching the same precision, which the authors attribute to the staged tightening of the reward region.","pith_inferences":["The paper stops at cumulative reward curves; a held-out pose-error table or physical grasp test would directly test the 'negligible errors' claim.","If the behavior transfers to a real depth camera, a single low-cost depth sensor plus this policy could form a complete suction-grasping front end, with no color camera and no hand-crafted vision pipeline.","The same graduated tightening of a reward region could be applied to other precision pose-regression tasks, such as peg insertion or tool alignment, where the end effector must converge to a tight tolerance."],"forward_implications":["A robot can solve this suction-grasping task with no RGB input, needing only a depth camera and a learned policy.","The curriculum schedule, with 19 lessons tightening position and yaw thresholds, is the component that produces precision, so reward shaping plus staged difficulty appears to be the key recipe.","DDPG and TRPO are not competitive on this task within the tested steps, suggesting that policy-gradient methods with staged curricula are better suited to continuous pose regression from images.","Sparse rewards do not generate learning here, so dense, continuously informative rewards are necessary for this kind of manipulation task."],"supporting_citations":[{"why":"Supplies the simulation platform and visual encoder that the setup modifies for depth input.","marker":"[1]"},{"why":"Provides the comparative baseline of off-policy deep reinforcement learning methods for vision-based grasping that this work extends to suction and 2.5D input.","marker":"[5]"},{"why":"Supplies the DDPG algorithm used as one of the compared policies.","marker":"[12]"},{"why":"Supplies the TRPO algorithm used as a comparison, which the paper finds unstable.","marker":"[15]"},{"why":"Supplies the PPO algorithm that, when combined with curriculum learning, yields the best results in the paper.","marker":"[16]"},{"why":"Supplies the curriculum learning strategy that progressively tightens the task difficulty.","marker":"[17]"}],"fun_headline_variants":["Deep RL learns suction grasps from 2.5D images alone","PPO with curriculum beats TRPO and DDPG on depth-only grasping","Single depth image guides PPO to precise suction grasps","2.5D vision only: PPO solves suction grasping in simulation","Depth-only DRL: PPO regresses grasp pose with negligible error"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reward thresholds (tooltip within 0.1 m in x and y, 0.01-0.02 m in z, and within 10 degrees of the block yaw, tightened to 0.01 m and 2 degrees in the final lesson) are treated as the definition of a successful grasp, but the paper never measures whether a tool configuration that meets them actually lifts the block.","fun_headline_variants_meta":{"raw":{"variants":["Deep RL learns suction grasps from 2.5D images alone","PPO with curriculum beats TRPO and DDPG on depth-only grasping","Single depth image guides PPO to precise suction grasps","2.5D vision only: PPO solves suction grasping in simulation","Depth-only DRL: PPO regresses grasp pose with negligible error"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000261,"raw_usage":{"total_tokens":1563,"prompt_tokens":884,"completion_tokens":679,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":500,"completion_tokens_details":{"reasoning_tokens":585}},"tokens_in":500,"tokens_out":679,"duration_ms":6828,"temperature":1.0,"reasoning_tokens":585,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:28:07.499230+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate the trained PPO-plus-curriculum policy on a held-out set of random block positions, orientations, and shapes, recording the physical outcome of the suction attempt (whether the block lifts and stays attached) together with the measured tooltip pose error. If the policy frequently enters the rewarded tolerance zone without achieving a grasp, the dense-reward proxy is the weak link; if it misses the zone, the pose-regression claim fails.","supporting_citations":[{"cited_title":"Deep reinforcement learning for vision-based robotic grasping: A simulated comparative evaluation of off- policy methods,","cited_arxiv_id":null,"evidence_quote":"Provides the comparative baseline of off-policy deep reinforcement learning methods for vision-based grasping that this work extends to suction and 2.5D input."},{"cited_title":"Continuous control with deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the DDPG algorithm used as one of the compared policies."},{"cited_title":"Cur- riculum learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the curriculum learning strategy that progressively tightens the task difficulty."}],"review_version":1}