{"id":"a240026f-eb7c-4257-a5f9-35193c247297","arxiv_id":"2509.06201","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A value-guided MPC policy trained on 2 million synthetic trajectories improves closed-loop 6-DoF grasping in clutter and adapts to object perturbations.","lead":"This paper presents a robot grasping controller that uses a learned value function inside model predictive control to adjust the gripper in real time while grasping. It reports higher success rates than open-loop and imitation baselines in simulated and real cluttered scenes.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never ablates the learned value function against the same MPC with a geometric distance-to-grasp cost, so the reported gains may be due to closed-loop MPC rather than the learned value function.","rationale":"The reader's weakest assumption is that the training labels are not physics-validated; that is a real limitation and is acknowledged by the authors in Section VII and Appendix D. However, the full system is evaluated in Isaac Sim with physics and in the real world, so the proxy-label concern is partially mitigated empirically. The more direct threat to the paper's central contribution is the missing ablation of the learned value function against a geometric grasp cost. The paper claims geometric costs are inadequate and that the learned value function is what makes Grasp-MPC robust, but no experiment compares Grasp-MPC to an MPC variant with a distance-to-grasp cost. This is not an external-consensus dispute; it is an internal experimental gap: the component that defines the method's novelty is never isolated. The existing CONDITIONAL verdict is appropriate, but the condition should include this ablation. The Figure 5/text inconsistency (74.9% vs 73.6%) and absence of code/data further reduce confidence, but the missing value-function ablation is the load-bearing issue.","tokens_in":17489,"tokens_out":12466,"duration_ms":159249,"concrete_test":"Run an 'MPC-Geom' baseline with the same CuRobo/MPPI hyperparameters and cost terms, but replace the value-based grasp cost in Eq. 5 with a squared SE(3) distance from the current end-effector pose to the target grasp pose (the same ground-truth, noisy, or M2T2 target used for each trial). Evaluate on the same 5,400 FetchBench cases and, if feasible, the real-world scenes. If MPC-Geom is within 3 percentage points of Grasp-MPC on the M2T2 split, the learned value function is not the load-bearing component and the central claim fails; if it is 5+ points worse, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"In Section IV-C the grasp cost is the learned value function V_theta (Eq. 5), combined with CuRobo costs in Eq. 6. The central novelty is that this learned value, rather than a geometric distance cost, enables robust closed-loop grasping. Yet no experiment replaces V_theta with a distance-to-target-grasp-pose cost inside the same MPPI/CuRobo framework. The comparisons in Section V and VI are against OSC (open-loop linear execution) and end-to-end IL/RL policies; none of these isolate the value function. The paper asserts in Section I that 'traditional geometric cost functions based on distances to predicted grasp poses are sensitive to prediction errors and fail to exploit MPC's full closed-loop capability,' but this is never tested. Therefore the headline results (74.9% GT, 60.3% noisy, 67.2% M2T2; real-world 74.4% vs 41.1%) could be explained entirely by closed-loop MPPI tracking a pre-grasp-to-grasp target with collision avoidance, with the learned value function contributing little. Without this ablation, the load-bearing role of the 2M-trajectory value function is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Grasp-MPC, a closed-loop 6-DoF visual grasping framework that combines model predictive control with a learned value function. A value function is trained on a synthetic dataset of about 2.1M trajectories generated by CuRobo motion planning from perturbed pre-grasp poses to annotated grasp poses across 8,515 Objaverse objects, with sparse success labels based on whether the motion plan reaches a feasible grasp pose. At deployment, an off-the-shelf grasp predictor (M2T2) proposes target poses, a motion planner drives the robot to a pre-grasp pose, and MPPI in CuRobo minimizes the learned value cost together with collision and smoothness costs. The method is evaluated in FetchBench-style simulation (5,400 problems) and in three real-world tabletop/shelf scenes (30 trials per condition), reporting grasp success of 74.9% with ground-truth poses, 60.3% under pose noise, 67.2% with M2T2 poses, and 74.4% overall in the real world, outperforming open-loop, diffusion policy, transformer policy, and IQL baselines.","tokens_in":17847,"tokens_out":5554,"duration_ms":69481,"significance":"If the reported results are fully supported, the paper makes a valuable empirical contribution: it demonstrates that a value function trained from large-scale synthetic motion-planned trajectories can serve as an MPC task cost for closed-loop grasping in clutter, with real-world validation on novel objects. The scale of the dataset (2M trajectories, 115M states) and the use of both successful and failed trajectories are strengths, as are the modular integration with collision avoidance and the evaluation across multiple scene types. The central novelty, however, is the learned value cost, and the paper does not currently isolate this contribution from the closed-loop MPC mechanics. The simulation results are extensive, but the real-world claims rest on small samples without uncertainty quantification. The label definition used to train the value function is also a correctness risk that deserves direct scrutiny. The overall idea is promising and worth pursuing, but the evidence as presented is not yet conclusive for the strongest claims.","major_comments":[{"comment":"The central contribution is the learned value function as the grasp cost in MPC, yet there is no ablation that replaces V_theta with a geometric distance-to-grasp cost within the same MPPI/CuRobo framework. The comparisons in Section V are against OSC (open-loop linear execution), Transformer/DP policies, and IQL; none of these isolates the value function. The claim in Section I that geometric distance costs are sensitive to prediction errors and fail to exploit MPC's closed-loop capability is therefore never directly tested. It is plausible that closed-loop MPPI with collision avoidance and smoothness costs, tracking the same pre-grasp-to-grasp motion, would produce most of the reported gains over OSC, with the learned value contributing little. Please add an ablation that uses the same MPC and hyperparameters but replaces Eq. (5) with a cost such as ||T_ee - T_grasp||^2 (or a similarly","section":"§IV-C, Eq. (5)-(6); §V-B and §V-C"},{"comment":"The training labels define a trajectory as successful when the motion plan reaches the annotated grasp pose and the pose is marked feasible by the GraspGen/ACRONYM annotation process, not when the grasp is physically validated. The text states explicitly: 'We do not validate these trajectories with simulation to accelerate data collection.' This means the value function may be learning motion-planning reachability to an annotated pose rather than physical grasp success (e.g., contact stability, slip, gripper-object interference). The real-world results mitigate this concern for the evaluated scenes, but they do not disambiguate whether the value function's guidance is superior because it reaches physically good poses or because it simply tracks planned reachable poses. For a stronger claim, please provide evidence on the correlation between the motion-planning-based labels and physical g","section":"§IV-A, Eq. (1)"},{"comment":"The real-world experiments use only 30 trials per scene condition (5 objects x 3 poses x 2 trials per method) and report point estimates without confidence intervals or significance tests. For example, the empty-table condition is 50.0% vs 86.7%, the cluttered-table 43.3% vs 67.7%, and the shelf 30.0% vs 70.0%; with n=30 per condition, these differences have very wide confidence intervals. The '33.3% improvement' headline in the abstract is therefore not statistically well grounded. Please report per-trial results, confidence intervals (e.g., Wilson intervals), and, where possible, a significance test or effect-size measure. This is load-bearing for the real-world claims, which are a key part of the paper's contribution.","section":"§VI-B and Fig. 8"}],"minor_comments":[{"comment":"The text says 'Grasp-MPC (73.6%)' when comparing with IQL, but Figure 5 and the subsequent text report 74.9% for the ground-truth condition. One of these is a typo and should be corrected.","section":"§V-B"},{"comment":"The translation noise is written as 'U(−0.04cm,0.04cm)'. If this is intended to be meters or centimeters, the notation is inconsistent with the surrounding numerical values (0.04 m would be 4 cm; 0.04 cm is 0.4 mm). Please clarify the units.","section":"§IV-A"},{"comment":"The caption's numeric list is not visually keyed to the bars in a way that is easy to parse. Please add value labels directly above each bar or restructure the caption by group (ground-truth, noisy, M2T2) for readability.","section":"Figure 5 caption"},{"comment":"There are typos in the appendix text: 'retreive pose' and 'consistenly' should be 'retrieve' and 'consistently'. Also, reference [19] and [44] both cite the same MPPI paper, and [49] and [55] both cite Grounding DINO; duplicate references should be consolidated.","section":"Appendix D"},{"comment":"The evaluation metric excludes trials where motion planning to the pre-grasp pose failed, and Appendix D shows this reachability varies by method (e.g., ~84-87%). The paper should state more prominently that the headline success rates are conditional on successful pre-grasp reaching, since the absolute task-success rate is lower for all methods.","section":"§V-A"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline for this one: a large, serious empirical effort on closed-loop grasping that achieves real gains, but the paper never isolates the thing it claims is new. The authors train a point-cloud-conditioned value function on 2M synthetic trajectories and plug it into CuRobo's MPPI. On FetchBench they get 74.9% success with ground-truth grasp poses versus 79.0% for an open-loop oracle, and in real-world cluttered scenes they get 74.4% versus 41.1% for an open-loop baseline. That is a meaningful empirical step, and the scale of the dataset and evaluation is genuinely impressive.\n\nWhat is actually new is the combination: a value function learned from hundreds of thousands of successful and failed grasp trajectories, conditioned on partial point clouds, used as a cost in a sampling-based MPC with collision avoidance. I haven't seen that exact recipe at this scale before. The authors also deserve credit for discussing the biggest limitation of their data pipeline - that trajectories are labeled by motion-planning success to a feasible grasp pose, not physics-validated grasps (Section IV-A) - and for being honest that physics simulation would probably help.\n\nThe soft spots match those concerns. The most serious is the missing ablation. The authors argue in the intro that geometric distance costs are sensitive to prediction errors, but they never run that baseline inside their own MPC framework. So we don't know whether the gains come from the learned value function or simply from closed-loop MPPI tracking a pre-grasp target with collision checking. That is load-bearing for the central claim. Second, the real-world evaluation is 30 trials per condition with no error bars, and the text says 73.6% where Figure 5 shows 74.9% - a small but real inconsistency. Third, no code or data are released, which limits reproducibility.\n\nMy overall read: this is a solid systems paper that deserves a serious referee, but the missing ablation should be requested before publication. If the authors can show that replacing the learned value with a simple distance cost degrades performance, the contribution becomes much more convincing. As it stands, I'd treat the result as a strong empirical recipe, not a clean validation of the value-guided approach.","headline":"Value-guided MPC for grasping is an impressive empirical system, but the missing ablation against a geometric cost means the central contribution is not yet isolated.","tokens_in":18287,"tokens_out":3575,"would_cite":true,"duration_ms":38206,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A learned value function makes robot grasping reactive in clutter","keywords":["model predictive control","grasping","value function","closed-loop manipulation","point cloud","offline reinforcement learning","motion planning","cluttered scenes"],"falsifier":"Relabel the same 2M trajectories using a physics simulation that actually closes the gripper and lifts the object, retrain the value function, and rerun the simulated and real evaluations. If grasp success does not improve—or if the physics-labeled value function performs no better than the planner-labeled one—then the planner-success proxy is not the limiting factor; if it improves substantially, the reported gains depend on that proxy rather than on the MPC loop itself.","tokens_in":17469,"feed_emoji":"🤖","tokens_out":10828,"duration_ms":118202,"temperature":0.7,"pith_summary":"Grasp-MPC is a closed-loop 6-DoF grasping policy: instead of committing to a precomputed grasp trajectory, the robot re-plans every control step using a learned value function as the task cost. The value function is trained on 2.1 million synthetic reach-to-grasp trajectories spanning 8,515 objects, with sparse labels that mark whether a motion plan reached an annotated feasible grasp. Because the value function consumes a segmented point cloud of the target object and the gripper pose relative to its center, it transfers to new objects and to cluttered scenes without retraining. In simulation it nearly matches an open-loop oracle when grasp poses are perfect (74.9% vs 79.0%) and beats all baselines when poses are noisy or predicted (60.3% and 67.2%); in real cluttered scenes it reaches 74.4% vs 41.1% for the open-loop pipeline. The paper's point is that the main failure mode of open-loop grasping—grasp-prediction error and object movement—can be handled by control rather than by better prediction alone.","feed_headline":"A learned value function makes robot grasping reactive in clutter","feed_subtitle":"Trained on 2M synthetic trajectories, it absorbs grasp-pose errors and beats open-loop baselines in clutter.","key_machinery":"The load-bearing object is the learned grasp-value function V(x_t), trained with the Bellman error objective on 2M synthetic trajectories. Its input is a centered segmented point cloud of the target object, encoded by a point-cloud network, concatenated with the end-effector pose relative to the point-cloud centroid; its output is a positive scalar approximating expected discounted cost-to-go. In the MPC loop this value is summed over the horizon as the task cost, with weight 1000, alongside the planner's default world-collision, self-collision, and smoothness costs. The value function is what gives MPC a notion of 'this trajectory ends in a good grasp' beyond geometric distance to a possibl","core_discovery":"The central claim is that a value function trained offline on motion-planned reach-to-grasp trajectories can serve as the task cost inside model predictive control, and that this makes closed-loop visual grasping work for novel objects in clutter. The value function predicts expected discounted cost-to-go from a segmented target point cloud and the gripper pose relative to the object centroid, trained by Bellman-error minimization with sparse costs that are zero only near a feasible grasp goal. At deployment, MPC minimizes the planner's default costs plus a discounted sum of this learned value over a 30-step horizon, so it steers toward states that predict grasp success rather than toward a","pith_inferences":["A testable extension the paper does not run: use the trained value function to score and select among candidate grasp poses before execution, rather than trusting the prediction model's ranking; the same value function could serve as a grasp-quality filter.","The phase analysis hints that grasps chosen by value-guided MPC survive a subsequent retrieve motion more often (99.8% vs 87.5% with predicted poses); if that holds, the value function implicitly encodes grasp stability and could be reused as a post-grasp transport-quality predictor.","Because the training labels never check physics, the reported numbers are likely a lower bound on what the same pipeline could achieve; relabeling the 2M trajectories with a quick physics grasp check is the cheapest high-value follow-up.","The moving-object experiment (60% success on 30 trials) is promising but small; scaling perturbations and using continuous object motion would test whether the value function degrades gracefully or fails abruptly under distribution shift."],"forward_implications":["Closed-loop grasping can recover from grasp-pose errors up to ±2 cm and ±18°, the regime that breaks open-loop methods; this directly targets the stated main failure mode of grasping in clutter.","Because MPC itself is the policy, no behavioral policy needs to be extracted from the offline data; the value function can be reused with any sampling-based MPC and any added constraint cost.","The value function transfers from synthetic empty scenes to real cluttered scenes without physics simulation in the training labels, so large-scale data generation via motion planning is sufficient for generalization.","Since labels are only success/failure, the value function can be fine-tuned with real-world demonstrations or with physics-simulated relabeling without changing the architecture or the MPC loop.","The approach is not limited to grasping: the same value-function-plus-MPC pattern applies to any manipulation task for which demonstrations can be generated by motion planning."],"supporting_citations":[{"why":"Supplies the off-the-shelf grasp-prediction model whose noisy outputs are the closed-loop policy's main challenge in simulation and real deployment.","marker":"[1]"},{"why":"Provides the 8,515 objects used to generate the 2M synthetic grasp trajectories.","marker":"[9]"},{"why":"Defines the antipodal grasp annotation style (feasible/infeasible grasps) that the dataset inherits.","marker":"[12]"},{"why":"Provides the simulated benchmark environments (adapted to a different simulator) used for the 5,400-problem evaluation.","marker":"[14]"},{"why":"Provides the grasp pose annotations used to define pre-grasp and goal poses in data generation.","marker":"[39]"},{"why":"Supplies the motion planner used to generate all training trajectories and the MPC library used for deployment.","marker":"[40]"},{"why":"Supplies the Bellman-error/temporal-difference objective used to train the value function.","marker":"[41]"},{"why":"Supplies the sampling-based MPC algorithm used as the closed-loop controller.","marker":"[44]"},{"why":"Supplies the point-cloud encoding architecture used in the value function.","marker":"[45]"},{"why":"Provides the diffusion policy baseline that Grasp-MPC is compared against.","marker":"[47]"}],"fun_headline_variants":["Value-guided MPC boosts grasping in clutter","Reactive grasping in clutter via value-guided MPC","Learned value guides MPC grasping in clutter","Value-guided MPC enables grasping in clutter"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing assumption is that a trajectory should be labeled a successful grasp if motion planning reaches a grasp pose annotated as feasible and the gripper joints are within 5 mm of the goal—without ever checking in physics simulation whether the gripper actually lifts the object.","fun_headline_variants_meta":{"raw":{"variants":["Value-guided MPC boosts grasping in clutter","Reactive grasping in clutter via value-guided MPC","Learned value guides MPC grasping in clutter","Value-guided MPC enables grasping in clutter"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001058,"raw_usage":{"total_tokens":4275,"prompt_tokens":742,"completion_tokens":3533,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":3488}},"tokens_in":486,"tokens_out":3533,"duration_ms":28967,"temperature":1.0,"reasoning_tokens":3488,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T23:57:00.874645+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Relabel the same 2M trajectories using a physics simulation that actually closes the gripper and lifts the object, retrain the value function, and rerun the simulated and real evaluations. If grasp success does not improve—or if the physics-labeled value function performs no better than the planner-labeled one—then the planner-success proxy is not the limiting factor; if it improves substantially, the reported gains depend on that proxy rather than on the MPC loop itself.","supporting_citations":[{"cited_title":"M2t2: Multi-task masked transformer for object-centric pick and place,","cited_arxiv_id":null,"evidence_quote":"Supplies the off-the-shelf grasp-prediction model whose noisy outputs are the closed-loop policy's main challenge in simulation and real deployment."},{"cited_title":"Objaverse-xl: A universe of 10m+ 3d objects,","cited_arxiv_id":null,"evidence_quote":"Provides the 8,515 objects used to generate the 2M synthetic grasp trajectories."},{"cited_title":"Acronym: A large-scale grasp dataset based on simulation,","cited_arxiv_id":null,"evidence_quote":"Defines the antipodal grasp annotation style (feasible/infeasible grasps) that the dataset inherits."},{"cited_title":"Fetchbench: A simulation benchmark for robot fetching,","cited_arxiv_id":null,"evidence_quote":"Provides the simulated benchmark environments (adapted to a different simulator) used for the 5,400-problem evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Bellman-error/temporal-difference objective used to train the value function."},{"cited_title":"Aggressive driving with model predictive path integral control,","cited_arxiv_id":null,"evidence_quote":"Supplies the sampling-based MPC algorithm used as the closed-loop controller."}],"review_version":1}