{"id":"75224418-b2d3-4d2d-afe9-4c08c6145254","arxiv_id":"2412.20451","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Adding a visual and textual chain-of-affordance reasoning step to a vision-language-action model improves robot manipulation success rates and generalization in the paper's evaluations.","lead":"CoA-VLA teaches a robot model to first reason through four types of affordances (what to grab, where to grab, where to put, and how to move) before predicting actions, using both text and images. The paper claims this improves performance over models like OpenVLA on real and simulated robot tasks, with better handling of new object poses and obstacles.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-robot evaluation is underpowered: 77 total trials, 5 more successes than DiffusionVLA, no error bars or significance testing, so the claimed 14.29% improvement is not statistically established.","rationale":"The reader's verdict is CONDITIONAL and cites small real-robot samples, cited baselines, and the annotation pipeline as concerns. I agree that the annotation pipeline is important for the mechanistic interpretation, but the most load-bearing issue for the paper's central empirical claim is statistical: the real-robot success-rate differences are computed from 11 trials per task, and the primary claimed advantage over the authors' own DiffusionVLA backbone (64/77 vs 59/77) is within the noise band of such small samples. If the real-robot effect is not real, the abstract's specific '14.29%' claim fails, and the paper's claim to outperform SOTA in real-world tasks is weakened. The LIBERO results offer more statistical power, but they rest on cited baselines and an ambiguous filtering sentence in Section 6.4, so they cannot fully compensate. My concern does not shift the overall verdict: the paper's contribution remains plausible and merits conditional acceptance pending stronger evidence. I therefore keep CONDITIONAL, while highlighting that the condition should include a properly powered real-robot evaluation and a clarification of the filtering procedure. This is a partial agreement with the reader: the reader noted the small samples but chose the annotation pipeline as the weakest assumption; I judge the statistical power of the real-robot evaluation to be more directly load-bearing for the central claim.","tokens_in":18315,"tokens_out":10775,"duration_ms":101302,"concrete_test":"Run each of the seven real-robot tasks with at least 30 trials per method under identical randomization and report per-task success counts with 95% binomial confidence intervals and a two-sided test (e.g., Fisher's exact or bootstrap difference of proportions) on the pooled 210+ trials. If the 95% CI for the difference between CoA-VLA and DiffusionVLA includes zero, the claimed 14.29% improvement is not established. Separately, clarify the 'We filter out the failure data' sentence in Section 6.4: if it applies to evaluation episodes rather than training demonstrations, recompute all LIBERO success rates without any filtering and re-check the comparison against OpenVLA.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 1 reports per-task n=11 and pooled totals of 64/77 for CoA-VLA vs 59/77 for DiffusionVLA. The difference is 5 successes, concentrated in three tasks (PourTea +2, PlaceBread +3, WipeWater +3); two tasks show no gain and NailHammer shows a loss. For n=77, the standard error of the difference in proportions is about 0.065, making the observed 8.5 percentage-point gap (14.29% relative) non-significant (roughly p>0.15). No confidence intervals or repeated runs are provided. The visual-generalization comparison (36/63 vs 28/63) is also marginal (p approx 0.08). Thus the abstract's headline claim that CoA-VLA 'exceeds its own diffusion backbone DiffusionVLA by 14.29% on real-robot tasks' is not supported by the data. The LIBERO benchmark provides more trials but relies on cited baseline numbers, and the supplementary Section 6.4 contains an ambiguous sentence 'We filter out the failure data' that, if applied to evaluation rollouts, would invalidate the success-rate computation. The central claim's real-robot component therefore rests on statistically underpowered evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CoA-VLA, a vision-language-action model that generates a chain of four affordances—object, grasp, spatial, and movement—in both textual and visual formats before predicting actions. The affordances are produced through a dynamic selection mechanism and injected into a diffusion-based policy via a visual-textual co-injection module. The authors evaluate the method on seven real-robot tasks and on the LIBERO benchmark, reporting improvements over OpenVLA, Octo, Diffusion Policy, and the DiffusionVLA backbone, as well as qualitative generalization results for unseen poses, free-space placement, and obstacle avoidance.","tokens_in":18554,"tokens_out":3466,"duration_ms":34982,"significance":"The paper addresses a timely and important question: whether explicit, structured affordance reasoning can improve VLA policies. The proposed four-part affordance taxonomy and the visual-textual co-injection module are well motivated, and the dynamic affordance selection is a sensible efficiency mechanism. The LIBERO evaluation uses 1500 trials with standard errors and three-seed averaging, and the ablations in Tables 4 and 5 provide useful evidence about the contribution of each component. The real-robot study covers seven tasks and includes both in-distribution and visual-generalization settings. If the results were rigorously established, CoA-VLA would be a meaningful advance in reasoning-augmented robot manipulation.","major_comments":[{"comment":"The headline real-robot claim is not statistically supported. In the in-distribution setting, CoA-VLA succeeds in 64/77 trials versus 59/77 for DiffusionVLA, a difference of only 5 successes concentrated in three tasks (PourTea +2, PlaceBread +3, WipeWater +3) with one task showing a loss (NailHammer -1). For n=77, the standard error of the difference in proportions is approximately 0.065, giving p>0.15; the visual-generalization comparison (36/63 vs 28/63) is also marginal (roughly p=0.08). No confidence intervals or significance tests are reported. Without additional trials or a more careful statistical analysis, the abstract's statement that CoA-VLA 'exceeds its own diffusion backbone DiffusionVLA by 14.29% on real-robot tasks' is not established. The qualitative claims in Section 5.3 about obstacle avoidance and free-space placement are based on very small numbers of scenarios and should be reported with explicit success counts.","section":"Section 5.1, Table 1"},{"comment":"The sentence 'We filter out the failure data and increase the image resolution to 224 x 224' is ambiguous and potentially damaging. If this filtering applies to evaluation rollouts, the reported success rates in Table 2 would be invalid. If it applies only to training-data cleaning or to the affordance-generation pipeline, that should be stated explicitly. The authors must clarify exactly where this filtering occurs and confirm that the same protocol is applied to all compared methods.","section":"Supplementary Section 6.4"},{"comment":"The automated affordance-annotation pipeline is a load-bearing component of the method, but its output quality is never quantified. The paper does not report any accuracy, agreement, or noise-level measurement for the GPT-4o, Grounding DINOv2/SAM, RoboPoint, and CoTracker labels, nor does it measure how often the model's self-generated affordances at test time are correct. Since the policy is trained on these labels and conditioned on them at inference, noisy or biased annotations could be the source of the observed gains rather than the proposed reasoning structure. I recommend adding an evaluation of affordance-prediction quality, an ablation with corrupted labels, or at minimum a manual inspection statistics for a held-out subset.","section":"Section 4.3"},{"comment":"The LIBERO results for Diffusion Policy, Octo, and OpenVLA are cited from the OpenVLA paper rather than re-run under the authors' exact protocol. Given that the authors also report using a different image resolution and a filtered dataset, the cross-paper comparison may not be apples-to-apples. The authors should either re-run the baselines with their own evaluation pipeline or provide evidence that the protocol differences do not affect the comparison.","section":"Section 6.4, Table 2"}],"minor_comments":[{"comment":"There is an apparent inconsistency: Section 5.1 states that DiffusionVLA is 'the same model our approach is built upon,' while Section 6.3 says the method is trained 'without relying on pre-trained weights from DiffusionVLA.' The authors should clarify whether the comparison is architectural or weight-initialization-based.","section":"Section 5.1 vs Section 6.3"},{"comment":"There are several typos, including 'Acknowledegments' in the acknowledgments heading, 'Wi U Water' in Section 6.2, and 'It can also represented' in Section 2. These should be corrected.","section":"General"},{"comment":"Reference [29] has a malformed author list that appears to concatenate many names without proper formatting; this should be fixed.","section":"References"},{"comment":"The inference-speed measurement is based on only 5 trials per task, which is a very small sample for a timing claim; adding more trials and reporting a variance estimate would strengthen the efficiency claim.","section":"Table 5"}],"recommendation":"major_revision","confidential_remarks":"The paper has a plausible and well-motivated contribution, but the real-robot evidence is statistically underpowered and at least one ambiguous sentence in the supplementary could call the reported results into question if it applies to evaluation. I would like to see the authors either collect more trials or soften the claims, clarify the filtering sentence, and provide quantitative validation of the affordance labels. The LIBERO results and ablations are the strongest part of the paper and offer a path to a defensible revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"CoA-VLA is a serious empirical systems paper. The new piece is the four-type affordance taxonomy (object, grasp, spatial, movement) rendered in both text and visual overlays, injected via a co-injection module into a diffusion VLA. That synthesis is not in the prior literature, though each ingredient is. The writing is clear and the method section gives enough detail to reproduce the pipeline if code and data were released (they are not).\n\nWhat the paper does well: LIBERO results include standard errors and 1500 trials, and the ablations show both textual and visual affordances matter, with dynamic selection cutting inference cost 6x without hurting success. The real-robot setup fine-tunes all baselines on the same data and uses the last checkpoint, which is the right way to compare.\n\nThe soft spots are real. The headline real-robot gain over DiffusionVLA is 64/77 vs 59/77—five successes. With n=77 that gap is not statistically significant (roughly p>0.15), and no error bars or repeated runs are given. The visual generalization comparison (36 vs 28) is also marginal. The abstract calls the 14.29% relative improvement a meaningful result; the data do not support that at conventional significance. On LIBERO, the baselines (Diffusion Policy, Octo, OpenVLA) are cited from OpenVLA's paper rather than re-run, which weakens the comparison. Most concerning is a line in Section 6.4: 'We filter out the failure data and increase the image resolution to 224x224.' If that means evaluation rollouts were filtered, the success rates are not valid. The authors need to clarify. The automated annotation pipeline (GPT-4o, Grounding DINOv2, SAM, RoboPoint, CoTracker) is never validated for accuracy, so we don't know how noisy the training signal is.\n\nI don't see a circularity or data-mining problem; the performance isn't reduced to a fitted parameter or to the authors' own prior results. The self-citation of DiffusionVLA is appropriate given it is the backbone.\n\nBottom line: the core idea is plausible and worth testing. The LIBERO evidence is decent, the real-robot evidence is underpowered, and that ambiguous filtering sentence must be resolved. A serious referee should see it, but the revision needs statistical rigor and a clearer evaluation protocol. I'd read the revision, not cite the current version.","headline":"A plausible affordance-reasoning synthesis for VLA policies, but the real-robot headline gain is five successes out of 77 trials and the evaluation protocol needs clarifying.","tokens_in":19087,"tokens_out":2158,"would_cite":false,"duration_ms":18471,"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 argues that a vision-language-action policy improves when it must first produce a chain of four affordances—what to manipulate and where, how to grasp, where to place, and how to move—before choosing an action.","keywords":["vision-language-action model","chain-of-affordance","affordance reasoning","robot manipulation","diffusion policy","visual prompting","generalization","multi-task learning"],"falsifier":"Take the trained model and, at test time, feed it deliberately wrong affordance overlays and texts; if success rates do not drop, the chain is not doing the causal work the paper attributes to it. A second decisive check is to compare the automatically generated affordance labels against human annotations on a small sample—if agreement is low, the training signal itself is unverified.","tokens_in":18148,"feed_emoji":"🤖","tokens_out":9341,"duration_ms":87624,"temperature":0.7,"pith_summary":"This paper is trying to show that a robot manipulation policy becomes more accurate, not by adding data or parameters, but by making the model produce a short reasoning chain before it acts. The chain consists of four robot affordances—which object to touch and where, which part of it to grasp, where to place it, and what collision-free path to follow—expressed both as text and as visual overlays on the camera image. This reasoning is injected into a diffusion-based action predictor, and the paper reports that the resulting model beats existing generalist robot policies on seven real-robot tasks, improves on its own backbone by 14.29 percent, and reaches a 79.8 percent success rate on the LIBERO simulation benchmark. A sympathetic reader would care because the recipe is cheap: it adds an intermediate reasoning stage to a vision-language-action model rather than scaling the model, and it appears to help precisely where robot policies usually fail—unseen object poses, cluttered scenes, and obstacle avoidance.","feed_headline":"Robot policies gain 14 percent with chain-of-affordance reasoning","feed_subtitle":"VLA that names object, grasp, placement, and path before acting beats larger baselines and generalizes to new scenes.","key_machinery":"The carrying mechanism is the chain-of-affordance: a fixed sequence of four affordance predictions interleaved as intermediate reasoning between the observation and the action. The visual-textual co-injection module is the other load-bearing piece: it takes text embeddings from the vision-language model and visual-affordance patch tokens from a small vision transformer, passes them through two transformer blocks, and injects the result into the diffusion action head with FiLM (feature-wise linear modulation) conditioning. Dynamic affordance selection, driven by proprioceptive state such as gripper and joint data, decides which affordances are needed at each timestep, keeping the added inference cheap. Together these pieces convert the model's own affordance guesses into explicit conditioning for action generation.","core_discovery":"The paper's central claim is that sequential affordance reasoning is a sufficient intermediate representation to improve vision-language-action policy learning. Before predicting an action, the model must first identify the target object and its location (object affordance), decide where to grasp it (grasp affordance), identify a collision-free placement region (spatial affordance), and trace a feasible motion path (movement affordance). The four affordances are generated by the model itself, rendered in both natural-language and pixel-aligned visual form, and fused into the action head. The discovery, as the paper reports it, is that this conditioning raises average success rates over the baselines it is compared with in both real-world and simulated tasks, and that the largest gaps appear during visual generalization, free-space placement, and obstacle avoidance.","pith_inferences":["Inference: part of the measured gain may come from the visual overlays acting as spatial priors rather than from the semantic chain; ablating overlays with shuffled coordinates would separate the two.","Inference: because the training labels are generated by an unmeasured automatic pipeline, the method's ceiling is tied to that pipeline; a small human-verified label set would quantify how much label noise remains.","Inference: the same chain could be generated once per subtask rather than per timestep and used by a slower high-level planner, which would reduce compute cost below even the paper's 6 Hz figure.","Inference: applying the chain to 3D or contact-rich affordances would be a natural extension, since 2D pixel points may under-specify grasps and placements in clutter."],"forward_implications":["Any vision-language-action model that can emit tokens and condition a diffusion head could adopt the chain-of-affordance pattern, since the additional module is small compared with the backbone.","Dynamic affordance selection keeps test-time reasoning cheap, making the method practical on a single GPU; the paper reports roughly 6 Hz inference versus 1 Hz when all affordances are forced.","The four-part taxonomy gives robot policies an inspectable intermediate output: failures can be attributed to the object, grasp, placement, or movement stage instead of the action decoder alone.","Explicit spatial and object affordances appear to reduce sensitivity to visual appearance, which is exactly the regime where the paper reports the largest gains over baselines."],"supporting_citations":[{"why":"It supplies the DiffusionVLA backbone that CoA-VLA extends and provides the direct comparison for the 14.29 percent real-robot gain.","marker":"[46]"},{"why":"It provides the primary baseline policy and the LIBERO evaluation setup that the simulation experiments follow.","marker":"[22]"},{"why":"It defines the LIBERO benchmark with four task suites used for the simulation results.","marker":"[30]"},{"why":"It supplies the large-scale robot dataset, filtered to 39K trajectories, used to pretrain the model on generated affordance data.","marker":"[21]"},{"why":"It supplies the diffusion policy baseline evaluated on the same real-robot tasks and LIBERO.","marker":"[10]"},{"why":"It supplies the generalist policy baseline compared against in the real-robot and simulation tables.","marker":"[36]"},{"why":"It contributes spatial-affordance point prediction used in the automated annotation pipeline.","marker":"[51]"},{"why":"It contributes the tracker used to record robot gripper trajectories as movement affordances in the annotation pipeline.","marker":"[19]"}],"fun_headline_variants":["Chain-of-affordance reasoning boosts robot policy success","VLA model reasons about grasp, place, path before acting","Affordance chains help robots generalize to new scenes","Plan grasp, place, and path before acting: robot wins","CoA-VLA: reasoning through affordances improves robot control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire result rests on the automatic annotation pipeline being accurate: if the object boxes, grasp points, placement points, and movement tracks it produces are noisy, the model is trained on corrupted reasoning, and the reported gains may come from the visual overlays acting as hints rather than from genuine affordance understanding.","fun_headline_variants_meta":{"raw":{"variants":["Chain-of-affordance reasoning boosts robot policy success","VLA model reasons about grasp, place, path before acting","Affordance chains help robots generalize to new scenes","Plan grasp, place, and path before acting: robot wins","CoA-VLA: reasoning through affordances improves robot control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00061,"raw_usage":{"total_tokens":2876,"prompt_tokens":1017,"completion_tokens":1859,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":633,"completion_tokens_details":{"reasoning_tokens":1777}},"tokens_in":633,"tokens_out":1859,"duration_ms":12901,"temperature":1.0,"reasoning_tokens":1777,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:21:35.489171+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained model and, at test time, feed it deliberately wrong affordance overlays and texts; if success rates do not drop, the chain is not doing the causal work the paper attributes to it. A second decisive check is to compare the automatically generated affordance labels against human annotations on a small sample—if agreement is low, the training signal itself is unverified.","supporting_citations":[{"cited_title":"DiffusionVLA: Scal- ing robot foundation models via unified diffusion and autore- gression","cited_arxiv_id":null,"evidence_quote":"It supplies the DiffusionVLA backbone that CoA-VLA extends and provides the direct comparison for the 14.29 percent real-robot gain."},{"cited_title":"OpenVLA: An open- source vision-language-action model","cited_arxiv_id":null,"evidence_quote":"It provides the primary baseline policy and the LIBERO evaluation setup that the simulation experiments follow."},{"cited_title":"LIBERO: Benchmarking knowl- edge transfer for lifelong robot learning","cited_arxiv_id":null,"evidence_quote":"It defines the LIBERO benchmark with four task suites used for the simulation results."},{"cited_title":"Octo: An open-source generalist robot policy","cited_arxiv_id":null,"evidence_quote":"It supplies the generalist policy baseline compared against in the real-robot and simulation tables."}],"review_version":1}