{"id":"900c9f7b-7ef7-47af-877a-719961deac8f","arxiv_id":"2504.19683","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"GPA-RAM fuses frozen grasp-pose features into an attention-Mamba imitation policy, reporting 87.5% average success on RLBench and 98%/56% on ALOHA cube transfer and bimanual insertion at about 71 FPS.","lead":"This paper combines a pretrained grasp detector with a fast Mamba-based network to make robot imitation learning more precise, and it reports state-of-the-art scores on standard benchmark tasks plus real-robot tests. The method, called GPA-RAM, lifts success on fine-grained insertion and stacking tasks while running at about 71 frames per second.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"GPA's claimed generic grasp prior is unsupported: the M2T2/ResNet detectors are pretrained on the same benchmarks used for evaluation, so the 2.6-point SOTA margin over ARP+ may be benchmark-specific leakage.","rationale":"The paper is a solid imitation-learning engineering effort with an honest ablation structure, multi-seed RLBench runs, portability checks on RVT2 and ACT, and real-robot trials. The central claim, however, is the SOTA margin over ARP+, and that margin rests entirely on GPA. The load-bearing premise is that GPA supplies a generic, transferable grasp prior; the paper does not establish this. Section IV.C names M2T2 and ResNet as the detectors but never states what data they were pretrained on. Since Section III.D says the detector is pretrained using grasp poses extracted from expert demonstrations—the same type of demonstrations used for policy training—the most plausible reading is that the detector has already seen the evaluation task distribution. That would make the 'augmentation' a form of benchmark-specific feature injection, not a general grasping prior, and would explain why RAM alone (84.7%) trails ARP+ (84.9%) while GPA-RAM (87.5%) surpasses it. The authors' own limitation paragraph, noting GPA hurts on 'Place Cups' and 'Push Buttons', reinforces that the effect is task-specific rather than a universal grasp affordance. A control experiment with a detector pretrained on disjoint data would settle this directly. The other issues (swapped BCE orientation in Eqs. 21-22, garbled Table I entries, missing Mamba-policy citations) are fixable reporting defects and do not threaten the central claim. Because the concern is addressable by disclosure and a control run, the reader's CONDITIONAL verdict is appropriate; no verdict change is needed.","tokens_in":18507,"tokens_out":4097,"duration_ms":39635,"concrete_test":"Run a control experiment where the M2T2 detector used in GPA is pretrained only on a disjoint set of RLBench tasks (or on a generic grasp dataset such as GraspNet-1Billion), never on the 18 evaluated tasks or their 100-demo training sets; retrain GPA-RAM and compare to ARP+ under identical evaluation. If the 2.6-point margin does not persist, the detector's benchmark-specific pretraining, not a generic grasp prior, is responsible for the SOTA claim. Ask the authors to release the detector's exact pretraining data and freeze/fine-tune configuration to interpret the result.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (87.5% vs ARP+ 84.9% on RLBench; 98%/56% vs 86%/16% on ALOHA) is carried almost entirely by GPA: RAM alone reaches 84.7%, below ARP+ (84.9%), and the ablation adds only +2.8%. GPA's sole novel input is F_GPA = phi_GPA(O_t, F_L, F_RAM) (Eq. 15), which concatenates features from a 'pre-trained grasp pose detector' (Section III.D). The paper says the detector is pretrained using grasp poses extracted from expert demonstrations, and Section IV.C states M2T2 is trained in a multi-task manner for RLBench and ResNet in a single-task manner for ALOHA—but never discloses the training data, the overlap with the 100-demo-per-task policy training sets, or whether the detector is frozen or fine-tuned ('can be frozen or fine-tuned', Section III.D). If the detector was pretrained on the same RLBench/ALOHA demonstrations (or the same task distribution) used for policy training and evaluation, then GPA injects task-specific knowledge rather than a transferable grasp prior, and the 'without additional grasp-pose data' framing is misleading. The authors' own limitation section concedes GPA hurts 'Place Cups' and 'Push Buttons', consistent with task-specific bias rather than a general grasping affordance. This enters at Section III.D (Eq. 15) and Section IV.C, and it directly determines whether the +2.6 point SOTA margin is a fair comparison or a confounded one.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes GPA-RAM, a framework for vision-based imitation learning in fine-grained robotic manipulation. The Grasp-Pretraining Augmentation (GPA) module injects features from a pretrained grasp-pose detector into an imitation policy through a fusion module called PLoFusion, with the stated goal of improving initial grasping without additional grasp-pose data or annotation. The Robotic Attention Mamba (RAM) architecture replaces Transformer encoders with a hybrid attention/state-space model to achieve real-time inference. The paper reports an average 87.5% success rate on 18 RLBench tasks, outperforming RVT2 (79.3%) and ARP+ (84.9%), and 98% and 56% success on ALOHA cube transfer and bimanual insertion, respectively, at approximately 71 FPS. It also presents real-robot experiments on UR5 and ARX R5 systems. Ablations compare GPA-RAM against RAM, C-RAM, GPA-C-RAM, GPA-RVT2, and GPA-ACT across RLBench and ALOHA.","tokens_in":18686,"tokens_out":7808,"duration_ms":80748,"significance":"If the headline results hold under a clean pretraining protocol, the paper would be a solid empirical contribution: the GPA module is architecture-agnostic, the RAM module addresses a real efficiency bottleneck, and the evaluation spans two sim benchmarks and two physical robot platforms. The measured inference-speed advantage over ACT (71 FPS vs 61 FPS) and the modular ablations are concrete strengths. The central SOTA claim is conditional, however, on two load-bearing issues: the pretrained grasp detector may have been trained on the same benchmark tasks used for evaluation, and the model-selection protocol appears to choose checkpoints on test episodes. These issues directly affect whether the 2.6-point margin over ARP+ on RLBench and the GPA-induced gains on ALOHA are due to transferable grasp priors or to benchmark-specific leakage and favorable selection.","major_comments":[{"comment":"The GPA feature F_GPA = phi_GPA(O_t, F_L, F_RAM) in Eq. (15) comes from a 'pre-trained grasp pose detector,' but the paper never discloses the detector's training data. Section IV.C states that M2T2 is used for RLBench and a single-task ResNet for ALOHA; M2T2 is a model trained on RLBench-style pick-and-place demonstrations, and the ResNet is trained on ALOHA demonstrations. If these detectors were pretrained on the same task distribution (or the same demonstrations) used for the 100-demo RLBench and 50-demo ALOHA policy training sets, then the GPA gains, including the +2.8% over RAM on RLBench and the +24% on ALOHA insertion, may reflect task-specific grasp knowledge rather than a transferable grasp prior. The paper must report: (i) the exact detector training split and its overlap with the policy training sets; (ii) whether the detector is frozen or fine-tuned in the main experiments (Section III.D says it 'can be frozen or fine-tuned,' without specifying); and (iii) an experiment with a grasp detector pretrained on object sets or tasks disjoint from the evaluation benchmark. Without this, the 'without additional grasp-pose data' framing is not established, and the SOTA margin over ARP+ is confounded.","section":"Section III.D and IV.C, Eq. (15)"},{"comment":"The implementation details state: 'Each task is trained on 100 demonstrations and tested four times on 25 unseen scenarios, using models from the final 30 training epochs. The best-performing model is selected for the final success rate comparison.' Selecting the checkpoint with the highest success on the test episodes is a form of test-set model selection, which inflates the reported success rate relative to a protocol that chooses a checkpoint on held-out validation episodes. Because the headline claim is the 2.6-point gap between GPA-RAM (87.5%) and ARP+ (84.9%), this selection procedure must be changed or justified. Please re-evaluate all methods under a validation-based early-stopping rule, or report the mean and standard deviation over all runs with an identical selection rule for every baseline and for all ablations.","section":"Section IV.C, model selection"},{"comment":"The paper does not quantify the uncertainty of the key comparisons. Given the reported protocol of four runs of 25 unseen scenarios per task, the per-task success rates are based on 100 episodes, and the paper reports no confidence intervals, standard errors, or hypothesis tests for the average success rates. The 2.6-point difference between GPA-RAM and ARP+ is small relative to the sampling noise from this protocol, and the model-selection issue in Section IV.C makes the point estimate additionally biased. Please report the number of independent training runs, the standard deviation across runs, the number of test episodes per task, and, where possible, paired comparisons across the 18 tasks so that the SOTA claim has a defensible statistical basis.","section":"Section IV.E and Table I"}],"minor_comments":[{"comment":"The state-space model equations are not written correctly: Eq. (7) appears to omit the time derivative and should be h'(t) = A h(t) + B x(t); Eqs. (9)-(10) also conflate the continuous-time matrix A with the discretized matrix. Please correct the notation.","section":"Eqs. (7)-(11)"},{"comment":"The column header 'Inference Time (fps)' is inconsistent with the reported values, which are frame rates, not times. Rename it to 'Inference Speed (FPS)' and specify whether the FPS is measured for the policy forward pass only or includes image rendering.","section":"Table II and Section IV.F"},{"comment":"The claim that GPA works 'without additional grasp-pose data or annotation' appears to conflict with Section III.D, where the detector is pretrained using grasp poses extracted from expert demonstrations. Clarify whether this means no new data collection beyond the existing demonstrations, or no additional labels of any kind.","section":"Abstract and Section III.D"},{"comment":"The module name 'PLoFusion' is introduced but never defined concretely; please provide the exact concatenation and attention operations, including feature dimensions, so that Eq. (15) is reproducible.","section":"Section III.D, Eq. (15)"},{"comment":"The selection rule differs between benchmarks: for RLBench, 'the best-performing model is selected' from the final 30 epochs, while for ALOHA, 'the best model during training is chosen.' Specify whether these procedures were applied identically to all baselines and ablations, and describe how the 50 test trials per ALOHA task are used for model selection.","section":"Section IV.C and Table VI"},{"comment":"The real-world results are based on only 10 trials per task; please report exact counts (for example, 7/10) and, if used, any confidence intervals, so the reader can assess the reliability of the real-robot comparisons.","section":"Tables III and IV"},{"comment":"The introduction reports a '26.0%' improvement on ALOHA bimanual continuous tasks, while the abstract reports 12% and 40% improvements for the two tasks; state the aggregation rule (presumably the mean of the two per-task deltas) so the numbers are unambiguous.","section":"Introduction, contributions"}],"recommendation":"major_revision","confidential_remarks":"The main risk is the same-benchmark pretraining of the grasp detector in GPA; if the authors cannot demonstrate a disjoint pretraining split or otherwise show that the detector provides a transferable grasp prior, the SOTA claim should be moderated to a modular ablative result. The checkpoint-selection protocol is also a serious concern for the headline comparison. I would suggest asking for a revised version that reports detector training details, validation-based model selection, and statistical uncertainty before the paper is accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a well-executed engineering paper with two real, if incremental, contributions. The GPA recipe—fusing features from a pre-trained grasp-pose detector into RVT2, ACT, or RAM—is new, and the portability ablations are the strongest evidence that the idea has legs. The RAM hybrid (attention + Mamba) also runs comfortably above the ALOHA control frequency. The reported numbers are plausible, but the headline SOTA claim is less solid than it looks.\n\nWhat is genuinely new: using a demonstration-pretrained grasp detector as a frozen feature source for downstream policy learning (Eq. 15) is a different lever from the usual attention and diffusion backbones. The coarse-to-fine RAM design and its continuous-action variant are also new combinations. The ablations on RVT2 and ACT (GPA-RVT2, GPA-ACT) make the architecture-agnostic claim credible.\n\nWhere it is soft: the 87.5% vs 84.9% SOTA margin on RLBench rests entirely on GPA. RAM alone is 84.7%, just below ARP+. GPA's own contribution is only +2.8%, which is within the range where evaluation choices matter. The paper never discloses the grasp detector's training data, its overlap with the 100-demo RLBench/ALOHA policy training sets, or whether the detector was frozen or fine-tuned. M2T2 is trained on RLBench grasp poses; if those poses come from the same demonstrations used for policy training, GPA is injecting task-specific knowledge, not a generic grasp prior. That does not invalidate the empirical gain, but it makes the SOTA margin a comparison between a policy plus a task-pretrained feature extractor and policies trained from scratch. That is a real confound. The stress-test note is fair, and I agree with it.\n\nThe authors' own limitation section says GPA hurts Place Cups and Push Buttons, and Table I shows those drops are 7 and 6 points, not slight. Other defects are minor: Eqs. 21-22 have the BCE labels swapped (pred and ground-truth reversed), the code link is only a project page, and the related work skips prior Mamba policies such as RoboMamba and Mamba Policy. Best-checkpoint selection over the final 30 epochs is optimistic but standard; it should be disclosed.\n\nBottom line: this deserves a serious referee. The direction is credible, the issues are fixable with disclosure and re-evaluation rather than new theory, and the paper has enough real content that a reviewer's time would not be wasted. The review should ask for detector training-data details, freeze/fine-tune status, and a fixed-checkpoint or averaged re-run on the RLBench evaluation. If those come back clean, the paper is a solid IEEE TMM contribution.","headline":"Solid imitation-learning engineering with a real but underspecified grasp-prior module; the SOTA headline is fragile because RAM alone sits below ARP+ and the detector's training data is undisclosed.","tokens_in":19499,"tokens_out":2946,"would_cite":true,"duration_ms":29895,"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":"GPA-RAM claims that injecting frozen grasp-detector features into robot policies plus an attention-Mamba backbone raises average manipulation success to 87.5% on RLBench, 98% on ALOHA cube transfer, and 56% on bimanual insertion while…","keywords":["robot learning","imitation learning","3D manipulation","multi-modal perception","grasp pretraining","state space models","bimanual manipulation","real-time robotic control"],"falsifier":"Retrain the grasp pose detector on demonstrations of a held-out set of objects and tasks, then run the full GPA-RAM evaluation on the original RLBench and ALOHA tasks. If the average success rate falls back to the RAM-only 84.7% level (or below the 84.9% ARP+ baseline), the grasp-pretraining augmentation is memorizing its own training benchmark rather than supplying a transferable grasping prior.","tokens_in":18079,"feed_emoji":"🤖","tokens_out":8166,"duration_ms":69161,"temperature":0.7,"pith_summary":"This paper claims that end-to-end imitation policies can be made substantially more precise at grasping by re-using the grasp features already latent in their own task demonstrations. The mechanism is a Grasp-Pretraining Augmentation (GPA) module that fuses features from a pretrained grasp pose detector into the policy's spatial features, together with a hybrid Robotic Attention Mamba (RAM) backbone that keeps the added capacity from destroying real-time speed. On the 18-task RLBench benchmark the combined system reports an 87.5% average success rate, beating RVT2 by 8.2 points and ARP+ by 2.6; on ALOHA it reports 98% on cube transfer and 56% on bimanual insertion, up 12 and 40 points over ACT, at roughly 71 frames per second.","feed_headline":"Grasp-pose pretraining lifts robot success to 87.5% on RLBench","feed_subtitle":"Attention-Mamba policy with grasp pretraining beats prior models on RLBench and ALOHA at 71 FPS.","key_machinery":"The load-bearing machinery is the pair (GPA, RAM). GPA, Grasp-Pretraining Augmentation, is a module built around a pretrained grasp pose detector (M2T2 on RLBench, ResNet on ALOHA) whose output head is removed and whose body remains frozen or fine-tunable; its grasping features are concatenated with RAM's spatially aligned features in the spatial dimension and fused through a self-attention block called Pre-trained Location Fusion (PLoFusion). RAM, Robotic Attention Mamba, is a hybrid backbone that interleaves single-view self-attention blocks with Mamba selective state-space layers, processing virtual multi-view RGB-D images in a coarse-to-fine heatmap pipeline for discrete keyframe prediction and in a one-stage form for continuous action generation. The Mamba layers provide linear-time global context aggregation while the attention blocks preserve local spatial alignment, which is what lets the combined model keep both precision and real-time throughput.","core_discovery":"The central claim, on the paper's own terms, is that grasp precision is a bottleneck in fine-grained manipulation and that it can be fixed without new annotations or a separate grasping stage. GPA takes a pretrained grasp pose detector, removes its prediction head, and uses the remaining feature extractor (frozen or fine-tuned) to produce grasping features that are concatenated with the policy's spatially aligned features and fused by self-attention; this supplies detailed affordance cues that deep Transformer layers tend to wash out. RAM, the companion architecture, combines single-view attention with Mamba selective state-space layers to aggregate global scene context in linear time and refines a coarse heatmap into a fine position, so the augmented policy stays fast enough for real-time control. The reported consequences are the benchmark numbers cited in the abstract plus real-robot demonstrations on UR5 and ARX R5 systems.","pith_inferences":["A testable extension not pursued in the paper: pretrain the grasp detector on demonstrations whose objects and scenes are disjoint from the RLBench evaluation tasks; if the gains persist, the augmentation is transferable, and if not, part of the +2.6-point margin over ARP+ is benchmark familiarity rather than a general grasping prior.","Another implication left implicit: since GPA helps most on tasks where initial grasp accuracy dominates (insertions, stacking, sorting), the same feature-fusion recipe should transfer to other policy families such as diffusion policies, where gripper-conditioned generation could absorb the grasp prior differently.","The slight regression on placement-flexibility tasks suggests a phase-dependent or confidence-weighted fusion of GPA features could remove the trade-off, turning the reported limitation into a design handle for future work."],"forward_implications":["An existing Transformer-based policy can be upgraded with GPA by a feature-concatenation change rather than by adding a separate grasping stage, gaining precision without new grasp-pose annotations.","Attention-Mamba hybrids can replace pure Transformer encoders in discrete and continuous manipulation policies, preserving or improving accuracy while avoiding quadratic sequence cost; RAM alone already beats RVT2 by 5.4 points on RLBench.","On ALOHA, the policy's 71 FPS throughput is fast enough for closed-loop servoing at 50 Hz, so the gains in bimanual insertion (56% versus 16% for ACT) are achieved with real-time execution rather than offline planning.","Because the same GPA recipe works with two different detector backbones (M2T2 on RLBench, ResNet on ALOHA), the benefit is not tied to a particular grasp detector architecture."],"supporting_citations":[{"why":"RVT2 is the main Transformer baseline on RLBench; it is both the comparison GPA-RAM must beat and the architecture used in the GPA-RVT2 ablation showing GPA adds 4.9 points.","marker":"[16]"},{"why":"ARP+ is the previous state-of-the-art on RLBench whose 84.9% average is the direct benchmark for the claimed 87.5% result.","marker":"[17]"},{"why":"ACT is the ALOHA baseline and host architecture; GPA and RAM are adapted into ACT for continuous actions, with reported gains of 12 and 40 percentage points.","marker":"[9]"},{"why":"RLBench supplies the 18-task discrete benchmark, keyframe conventions, demonstrations, and evaluation protocol.","marker":"[11]"},{"why":"Mamba provides the selective state-space layers that give RAM linear-time sequence modeling and the reported inference speed.","marker":"[20]"},{"why":"M2T2 is the pretrained grasp pose detector whose features GPA fuses into the RLBench policy, making it the source of the grasping priors.","marker":"[49]"},{"why":"ResNet is the grasp pose detector used in the ALOHA variant, supporting the claim that GPA works across detector architectures.","marker":"[50]"},{"why":"CLIP extracts the language and image features that RAM aligns in its attention stage, conditioning the policy on task descriptions.","marker":"[42]"}],"fun_headline_variants":["Grasp-pretrained Mamba: 87.5% RLBench, 71 FPS","Pretrained grasp cues boost robot success to 87.5% on RLBench","No extra grasp labels: GPA-RAM reaches 87.5% on RLBench","Grasp pretraining + Mamba speed: 87.5% RLBench at 71 FPS","Grasp-aware Mamba policy beats RVT2 by 8.2% on RLBench"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The gain depends on the pretrained grasp detector providing a generalizable grasp prior; the paper does not disclose whether the detector's training demonstrations overlap with the 100-demonstration task sets it is evaluated on, so if they do, the margin over RVT2 and ARP+ may reflect learned benchmark familiarity rather than transferable grasping skill.","fun_headline_variants_meta":{"raw":{"variants":["Grasp-pretrained Mamba: 87.5% RLBench, 71 FPS","Pretrained grasp cues boost robot success to 87.5% on RLBench","No extra grasp labels: GPA-RAM reaches 87.5% on RLBench","Grasp pretraining + Mamba speed: 87.5% RLBench at 71 FPS","Grasp-aware Mamba policy beats RVT2 by 8.2% on RLBench"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001183,"raw_usage":{"total_tokens":4920,"prompt_tokens":1014,"completion_tokens":3906,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":3784}},"tokens_in":630,"tokens_out":3906,"duration_ms":27075,"temperature":1.0,"reasoning_tokens":3784,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:48:06.524128+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the grasp pose detector on demonstrations of a held-out set of objects and tasks, then run the full GPA-RAM evaluation on the original RLBench and ALOHA tasks. If the average success rate falls back to the RAM-only 84.7% level (or below the 84.9% ARP+ baseline), the grasp-pretraining augmentation is memorizing its own training benchmark rather than supplying a transferable grasping prior.","supporting_citations":[{"cited_title":"Rvt2: Learning precise manipulation from few demonstrations,","cited_arxiv_id":null,"evidence_quote":"RVT2 is the main Transformer baseline on RLBench; it is both the comparison GPA-RAM must beat and the architecture used in the GPA-RVT2 ablation showing GPA adds 4.9 points."},{"cited_title":"Autoregres- sive action sequence learning for robotic manipulation,","cited_arxiv_id":null,"evidence_quote":"ARP+ is the previous state-of-the-art on RLBench whose 84.9% average is the direct benchmark for the claimed 87.5% result."},{"cited_title":"Mamba: Linear-time sequence modeling with selective state spaces,","cited_arxiv_id":null,"evidence_quote":"Mamba provides the selective state-space layers that give RAM linear-time sequence modeling and the reported inference speed."},{"cited_title":"M2t2: Multi-task masked transformer for object-centric pick and place,","cited_arxiv_id":null,"evidence_quote":"M2T2 is the pretrained grasp pose detector whose features GPA fuses into the RLBench policy, making it the source of the grasping priors."}],"review_version":1}