{"id":"813f0ef3-9dfe-44b8-9d51-c471dda3f0ee","arxiv_id":"2608.06907","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A time-aware, position-conditioned reinforcement learning policy enables a quadruped to intercept thrown balls with higher success than a velocity-tracking baseline in the tested setups.","lead":"The authors built a quadruped robot system that watches a ball in flight, predicts where and when it will land, and runs to catch it using a time-aware AI policy. In head-to-head trials it caught more balls than a standard velocity-tracking approach, though absolute success rates stayed modest.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-world success-rate comparison is not standardized: throws were placed in method-specific feasible zones, so the claimed advantage over the velocity-tracking baseline is not established for the full 2 m workspace.","rationale":"The reader's weakest_assumption identifies the most load-bearing concern: the real-world comparison uses method-specific throwing zones, making the success-rate advantage potentially an artifact of throw placement. This concern is central because the abstract's headline claim is a comparative success rate over a defined workspace, and the experiments do not enforce a common distribution over that workspace. The paper explicitly acknowledges this limitation, which is why the verdict is CONDITIONAL rather than an unconditional accept. I agree with the reader's assessment: the evidence supports a modest, feasibility-oriented advantage, but the central claim as stated requires a standardized benchmark to be fully validated. I did not identify a different single concern that is more load-bearing; the baseline being a hand-tuned P controller is a secondary issue, but the lack of a standardized throwing distribution is sufficient to question the comparative claim. The proposed concrete test directly addresses the gap by using a fixed grid, equal per-bin trial counts, and confidence intervals, which would settle whether the claimed advantage is real or an artifact of the experimental design.","tokens_in":11104,"tokens_out":8111,"duration_ms":76032,"concrete_test":"Design a pre-registered real-world benchmark using a fixed grid of landing positions (e.g., radii 0.25, 0.75, 1.5 m at eight headings) and flight times drawn uniformly from 0.8–1.2 s, with at least 20 trials per bin per method. Run both the proposed policy and the velocity-tracking baseline on the same throw schedule (same ball trajectories, alternating methods) and report per-bin catch success rates with Wilson binomial 95% confidence intervals, plus an overall rate weighted by the pre-registered distribution. If the proposed method's success rate remains higher under this standardized protocol, the abstract's claim holds; if not, the feasibility-oriented comparison in §4.3 is insufficient to support it.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the direct target-conditioned policy achieves a higher success rate than the velocity-tracking baseline for landing spots within 2 m and flight times of 0.8–1.2 s. Section 4.3 explicitly states that throws were placed 'near their feasible catching zones' for each method and that the comparison is 'feasibility-oriented rather than a fully standardized workspace-wide benchmark.' Because the two policies were evaluated on different throwing distributions, the higher per-bin success rates in Table 4 could reflect easier throws for the proposed method (e.g., front-lateral landing points) rather than a genuine policy advantage. The abstract's claim covers the entire 2 m workspace, but the experiments do not sample that space identically for both methods. Additionally, per-bin trial counts are not reported; with only 100 total trials per method spread across three distance bins, some bins may have very few samples, making the reported percentages unreliable. Without a standardized throwing protocol and per-bin confidence intervals, the comparative success-rate claim is not rigorously supported. The paper's own acknowledgment of this limitation is appropriate, but it means the current evidence is conditional rather than conclusive.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a framework for quadrupedal ball catching in which off-board cameras detect the ball and the robot, a Kalman filter predicts the ball's landing position and landing time, and a reinforcement-learned locomotion policy is conditioned directly on the position error and remaining time rather than on intermediate velocity commands. The policy is trained with a teacher-student distillation scheme in IsaacGym and deployed on a Unitree Go2. The authors compare their method against a velocity-tracking baseline in simulation (MuJoCo) and in real-world trials, and report higher catch success rates for their method in the real-world comparison. The paper also introduces a time-aware reward with a duration mask and observes a rotation-prioritized interception behavior. The manuscript is candid about several limitations, including different flight-time ranges between simulation and real trials, a non-standardized real-world throwing protocol, and the absence of ablations isolating time conditioning, distillation, reward terms, and prediction error.","tokens_in":11335,"tokens_out":4921,"duration_ms":47101,"significance":"If the central comparative claim were conclusively established, the paper would make a useful system-level contribution: it demonstrates a closed-loop pipeline that combines perception, prediction, and time-conditioned locomotion on real hardware, and it provides evidence that direct position-and-time conditioning can be a practical alternative to velocity tracking for time-constrained interception. The paper also explicitly frames its contributions as system-level rather than claiming a fundamentally new learning algorithm, which is appropriate. The manuscript deserves credit for its honest acknowledgment of limitations, particularly in Sections 4.2 and 4.3, where the authors state that the simulation and real-world evaluations are not strictly like-for-like and that the real-world comparison is feasibility-oriented. However, the same acknowledged limitations directly affect the abstract's headline claim: the claimed higher success rate for landing points within 2 meters and flight times of 0.8-1.2 seconds is supported only by a real-world comparison in which throws were placed near each method's feasible catching zone rather than from one shared distribution.","major_comments":[{"comment":"The real-world comparison does not use a shared throwing distribution for the two methods. The text states that throws were placed 'near their feasible catching zones' for each method and that the comparison is 'feasibility-oriented rather than a fully standardized workspace-wide benchmark.' Consequently, the higher strict catch rates reported in Table 4 could reflect easier throws for the proposed method (e.g., front-lateral landing points) rather than a genuine policy advantage. Since the abstract makes a claim about the full 2-meter workspace, the authors should either provide a standardized throwing protocol with the same distribution for both methods or explicitly restrict the claim to the feasibility-oriented setup. At minimum, per-bin trial counts and confidence intervals are needed to assess whether the observed gaps are meaningful.","section":"§4.3, Table 4"},{"comment":"The simulation and real-world evaluations use different flight-time ranges: the simulation uses 1.15-1.35 seconds, while the real-world trials use 0.8-1.2 seconds. In addition, the simulation success criterion counts a ball that bounces out of the basket as a success, whereas the real-world strict 'catch' criterion requires the ball to directly hit the basket bottom. These differences mean that the abstract's specified regime (0.8-1.2 seconds flight time) is only evaluated in the non-standardized real-world test, and statements about sim-to-real gaps or improved sim-to-real behavior may conflate different task definitions. The authors should align these definitions or clearly separate the claims made for each regime.","section":"§4.2 vs §4.3"},{"comment":"Per-bin trial counts are not reported, and with only 100 total trials per method spread across three distance bins, the reported percentages may be based on very small denominators. For example, in the strict catch row for the 1-2 m bin, the gap between 25.00% and 36.84% would not be significant with typical bin sizes, and even the larger gaps in other bins lose strength if the effective sample size is small. The authors should report the number of trials per bin and, ideally, bootstrap confidence intervals or a statistical test for the success-rate differences.","section":"Table 4"},{"comment":"The deployed velocity-tracking baseline is not identical to the simulation baseline: the real-world experiments use Kp=5.0 for distances below 0.5 m instead of the simulation-tuned Kp=2.5, and both methods receive a shared yaw-command shaping rule. While these changes are disclosed and the yaw shaping is applied to both methods, the baseline policy used in the real-world comparison is therefore not the same policy that was tuned and evaluated in the simulation. The authors should justify that these modifications do not bias the comparison, or provide results with an unmodified baseline, so that the real-world advantage can be attributed to the proposed conditioning rather than to a retuned baseline.","section":"§4.3"}],"minor_comments":[{"comment":"The abstract contains grammatical issues, such as 'this work presents a system-level contribution that completes real-time robotic interception system,' and would benefit from careful proofreading.","section":"Abstract"},{"comment":"The notation for the duration mask is confusing: the mask is defined as M(tgo, D) with a duration threshold D, but the reward uses D_rew without a clear definition, and the reward weight for r_pos_time in Table 1 is not included in Eq. (4). Please clarify the relationship between D, D_rew, and the reward weight.","section":"§3.2.2, Eq. (4)"},{"comment":"Because the real-world landing-point distributions are shown separately for the two methods and the throwing protocol intentionally differs, the figure should be accompanied by a clear statement of how the distributions were generated and how many throws fell into each distance bin.","section":"§4.3, Figure 8"},{"comment":"The conclusion acknowledges that the experiments do not isolate the effects of time conditioning, teacher-student distillation, reward terms, or prediction error. This is an important confound and should be stated earlier in the experimental sections, not only in the conclusion, so that readers can properly interpret the comparisons.","section":"§5"},{"comment":"The introductory paragraph contains citation formatting errors, for example '[ 1, 2], use perception-based methods to enhance...' which reads as an incomplete sentence. Please fix the sentence structure and citation placement.","section":"§1"}],"recommendation":"major_revision","confidential_remarks":"This is a system-level robotics paper whose main quantitative claim rests on a real-world comparison that the authors themselves describe as feasibility-oriented rather than standardized. The methodological core is sound and the paper is transparent about its limitations, but the abstract currently overstates the strength of the evidence. If the authors can provide a shared-distribution comparison with per-bin trial counts, or alternatively retitle and reframe the claim as a feasibility demonstration, the paper would be suitable for publication. The lack of ablations is a limitation but not, in my view, a blocking one for a system paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper builds and deploys a quadruped that catches tennis balls using a policy conditioned directly on predicted landing position and time-to-impact, with a time-gated reward. That conditioning interface is a real addition over position-only conditioning, and the integrated multi-camera plus Kalman plus sim-to-real loop is a working system. The authors are also unusually candid about several limitations, which I respect.\n\nWhere it gets soft: the abstract and Section 4.3 claim higher success rates across the 0-2 m workspace, but the real-world trials deliberately threw to each method's feasible catching zone. That makes the per-bin percentages in Table 4 hard to compare; the baseline may have faced systematically harder throws. Per-bin trial counts and confidence intervals are missing, and there are no significance tests. The simulation and real trials also used different flight-time ranges (1.15-1.35 s vs 0.8-1.2 s), so the 'smaller sim-to-real gap' statement is not actually supported by the numbers. These issues are acknowledged in the paper as feasibility-oriented, but the abstract does not carry that caveat. The baseline's kp was also changed during deployment for short distances; this is stated, but it further complicates the headline comparison.\n\nNone of this kills the contribution. The system works, the time-aware reward is a sensible idea, and the simulation experiments give some evidence that the policy beats velocity-tracking at short and long distances, even though the baseline wins the middle bin. The lack of a standardized throwing protocol is a fixable experimental flaw, not a conceptual one.\n\nThis paper is for legged-locomotion and robot-catching researchers, especially those working on time-constrained reaching. It deserves a serious referee. I would ask for revision with a standardized throwing protocol, per-bin trial counts, confidence intervals, and ideally release of video or code. The central claim is probably right in spirit, but the current evidence does not prove it as stated.","headline":"Solid system paper with a genuinely time-conditioned policy interface, but the real-world throwing protocol doesn't back the workspace-wide success-rate claim in the abstract.","tokens_in":11853,"tokens_out":2798,"would_cite":false,"duration_ms":27536,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a quadruped robot catches thrown balls more reliably when its reinforcement-learned policy is conditioned on the predicted landing point and time-to-impact rather than on velocity commands.","keywords":["quadruped locomotion","ball catching","time-constrained interception","reinforcement learning","sim-to-real transfer","vision-guided control","trajectory prediction","position-conditioned policy"],"falsifier":"Run both policies against one fixed set of throws sampled uniformly across the 0–2 m range with identical flight-time distribution and fixed robot start; if the target-conditioned policy does not beat the velocity-tracking baseline on that common distribution, the paper's central success-rate claim fails in the stricter sense.","tokens_in":10897,"feed_emoji":"🎾","tokens_out":10630,"duration_ms":92394,"temperature":0.7,"pith_summary":"Ball catching is a sharp test of quadruped agility: the robot must perceive a flying ball, predict where and when it will hit the ground, and get there in time. This paper claims that the right way to command that motion is not a velocity setpoint but the spatiotemporal target itself—the predicted landing position and the remaining time until impact. The authors assemble this into a complete system with off-board cameras, a Kalman filter for landing-point and time prediction, a teacher-student RL policy trained with time-gated rewards, and deployment on a Unitree Go2 with an onboard inference computer. In their simulation and real-world trials, the position-and-time-conditioned policy catches more balls than a velocity-tracking baseline, and it shows a smaller drop after sim-to-real transfer.","feed_headline":"Robot dog catches more balls when told where and when they land","feed_subtitle":"Telling a robot dog where and when a ball will land beats velocity commands in sim and real trials","key_machinery":"The load-bearing mechanism is a time-gated, position-conditioned reward that couples spatial accuracy with arrival time: $r_{\\mathrm{pos,time}} = M(t_{\\mathrm{go}}, D)\\, [1 + (\\lVert p_{\\mathrm{target}} - p_{\\mathrm{robot}}\\rVert / \\sigma_p)^2]^{-1}$, where the duration mask $M(t_{\\mathrm{go}}, D) = \\frac{1}{D}\\mathbb{I}(t_{\\mathrm{go}} \\le D)$ switches the reward on only when the remaining time is within the interception window. The policy's observation is the target position error $(\\Delta x, \\Delta y)$, orientation error $\\Delta \\theta$, normalized remaining time $t_t$, proprioception, and a short action history. Training uses concurrent teacher-student distillation in IsaacGym, with domain randomization, curriculum expansion of target distances, and RND exploration; the deployed student policy sees only the target commands and proprioception. The perception module supplies the predicted landing point and time through YOLO-based detection with an HSV fallback and a Kalman filter under a parabolic motion model.","core_discovery":"The paper's central claim is that conditioning a legged locomotion policy on the predicted landing position and time-to-impact—rather than on an intermediate velocity command—produces better dynamic interception for landing spots within 2 m and flight times of roughly 0.8–1.2 s. In simulation, the time-conditioned target-reaching policy reaches 86.73% catch success for targets within 0.5 m and 35.63% for 1–2 m, whereas the velocity-tracking baseline reaches 67.39% and 16.28% in those bins; the baseline is better in the middle band. In real trials on a Unitree Go2, the proposed policy's strict catch rates are 37.5%, 24.4%, and 36.8% across the three distance bins, against 6.25%, 9.62%, and 25.0% for the baseline, while its track (rim-contact) rates stay above 60%. The authors interpret these results as showing that explicit spatiotemporal conditioning, together with a time-gated reward, yields a rotation-prioritized, time-synchronized motion that degrades less under sim-to-real transfer. They are careful to describe the real comparison as feasibility-oriented, with throws placed near each method's feasible catching zone.","pith_inferences":["If time conditioning is the active ingredient, then removing the $t_t$ observation or the duration mask should measurably reduce catch rates; the paper does not run this ablation, so the causal role of the timing signal is an inference, not a demonstrated fact.","The same target-space interface—predicted position plus time-to-impact—could be applied to other dynamic tasks such as timed jumping, interception with a manipulator, or dodging, since the policy is not tied to ball-specific visual features.","The paper's own feasibility-oriented real-world protocol means the reported advantage is not yet a statement about general superiority; a uniform throw distribution across the whole workspace would be a stricter test.","The observed rotation-prioritized gait emerged without an explicit velocity-direction reward, suggesting that time-constrained RL can discover nontrivial motor strategies; whether this is robust across terrain and robot models remains open."],"forward_implications":["A quadruped can perform time-constrained interception with a roughly one-second horizon when commanded directly with predicted landing position and remaining time, instead of velocity setpoints.","Time-gated rewards produce a distinct motion strategy—turn first, then sprint—that outperforms lateral-sliding velocity tracking at short and long catch distances in the paper's simulation tests.","Explicit target-space commands transfer more cleanly to hardware: in real trials the strict catch rate of the proposed policy stays between 24% and 38% while the velocity-tracking baseline falls below 10% at short range.","Because the policy receives predicted landing point and time rather than live ball kinematics, brief detection dropouts are bridged by the Kalman filter without requiring the network to learn ballistic dynamics implicitly."],"supporting_citations":[{"why":"Concurrent teacher-student distillation framework used to train the deployable student locomotion policy.","marker":"[22]"},{"why":"Proximal Policy Optimization, the RL algorithm used to train the teacher policy.","marker":"[21]"},{"why":"IsaacGym parallel physics simulation used for policy training with domain randomization.","marker":"[23]"},{"why":"Source of the spatial-tracking reward formulation that the paper's time-aware reward terms follow and extend.","marker":"[14]"},{"why":"Introduces the duration mask M(t_go, D) used to gate rewards by remaining time.","marker":"[6]"},{"why":"YOLOv13n object detector used to localize the ball in the perception pipeline.","marker":"[17]"},{"why":"Position-conditioned end-to-end locomotion approach that the paper builds on for direct target commands.","marker":"[4]"}],"fun_headline_variants":["Direct target conditioning beats velocity commands for ball-catching robots","Predicting ball's landing point and time improves quadruped interception","Robot dog's new trick: aim for the predicted landing spot, not a velocity","Spatiotemporal conditioning boosts ball catch rates in sim and real trials","Quadruped robot catches better by targeting predicted landing point and time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The real-world comparison assumes that throwing the ball into each method's own feasible catching zone is a fair basis for comparing success rates; the paper states the trials were feasibility-oriented rather than a fully standardized workspace-wide benchmark.","fun_headline_variants_meta":{"raw":{"variants":["Direct target conditioning beats velocity commands for ball-catching robots","Predicting ball's landing point and time improves quadruped interception","Robot dog's new trick: aim for the predicted landing spot, not a velocity","Spatiotemporal conditioning boosts ball catch rates in sim and real trials","Quadruped robot catches better by targeting predicted landing point and time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000787,"raw_usage":{"total_tokens":3543,"prompt_tokens":1085,"completion_tokens":2458,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":701,"completion_tokens_details":{"reasoning_tokens":2367}},"tokens_in":701,"tokens_out":2458,"duration_ms":16153,"temperature":1.0,"reasoning_tokens":2367,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T18:33:18.891607+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run both policies against one fixed set of throws sampled uniformly across the 0–2 m range with identical flight-time distribution and fixed robot start; if the target-conditioned policy does not beat the velocity-tracking baseline on that common distribution, the paper's central success-rate claim fails in the stricter sense.","supporting_citations":[{"cited_title":"Cts: Concurrent teacher-student reinforcement learning for legged locomotion.IEEE Robotics and Automation Letters, 9:9191– 9198, 2024","cited_arxiv_id":null,"evidence_quote":"Concurrent teacher-student distillation framework used to train the deployable student locomotion policy."},{"cited_title":"Anymal parkour: Learning agile navigation for quadrupedal robots.Science Robotics, 9:eadi7566, 2024","cited_arxiv_id":null,"evidence_quote":"Source of the spatial-tracking reward formulation that the paper's time-aware reward terms follow and extend."},{"cited_title":"Advanced skills by learning locomotion and local navigation end-to-end","cited_arxiv_id":null,"evidence_quote":"Position-conditioned end-to-end locomotion approach that the paper builds on for direct target commands."}],"review_version":1}