{"id":"614c887b-c7e5-4f2a-ae68-d2e295b85ec9","arxiv_id":"2411.18562","paper_version":6,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A dual-phase diffusion planner with dynamics-consistency and LLM-written guidance achieves strong success on goal-adaptive dexterous manipulation in simulation.","lead":"This paper introduces a diffusion planner for dexterous robot hands that generates hand states, object states, and actions together, guided by contact-aware and dynamics-consistency objectives. The authors report success rates around 70 percent on goal-shifted simulated manipulation tasks, roughly doubling prior baselines on door opening with new target angles.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Unvalidated dynamics model in Eq. 12 is the most load-bearing assumption: it must extrapolate to OOD goal states, yet no error or ablation evidence is provided.","rationale":"The reader's identified weakest assumption is the same one I regard as most load-bearing: the learned dynamics model is the only component that explicitly couples states and actions, and its extrapolation behavior is untested. I considered the non-differentiable Heaviside penalties (Eqs. 14, 16) as an alternative; however, the appendix's LLM-generated code uses smooth L2 penalties and adaptive scaling, so the actual implementation almost certainly differs from the paper's formal description. That is a serious presentation flaw, but it does not directly threaten the measured success rates. The LLM-guidance component is also secondary because Table 5 shows it underperforms Human Craft, and the headline results use Human Craft. The dynamics model concern, by contrast, targets the mechanism the paper claims is responsible for the gains; if T extrapolates poorly, the reported improvements could be coming from the dual-phase interaction guidance alone, leaving the dynamics-based contribution unjustified. The proposed test would settle this by measuring OOD prediction error and ablating eps_dyn on the actual OOD tasks. Because the paper is currently a strong CONDITIONAL (needs validation and code), my read leaves the reader's verdict unchanged.","tokens_in":29671,"tokens_out":11785,"duration_ms":108812,"concrete_test":"Collect (s_t, a_t, s_{t+1}) triples from DexHandDiff's simulation rollouts on each OOD goal (Door 110, Close, Pen Half, Hammer Half, Relocate Half) and on Door 90 as reference. Compute T's mean normalized one-step prediction error against the simulator, split by pre/post-contact and by whether the state is in the demonstration support. If the OOD error is within a small factor (e.g., 2x) of the in-distribution error, the concern is resolved. If not, ablate eps_dyn by setting its scale to zero on those same OOD tasks; if success rates do not drop, the paper must either remove the claim that dynamics guidance is essential or provide a corrected validation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim rests on a pipeline whose only physics-consistency term is the learned dynamics model T(s,a) in Eq. 12, trained exclusively on demonstration data and used at inference as eps_dyn in both phases (Eq. 13). For the goal-adaptive evaluations (Door 110, Close Door, Pen left-hemisphere, Hammer half-drive, Relocate half-side), the generated trajectories contain states and actions outside the support of the demonstrations; T must extrapolate accurately for eps_dyn to act as a correct penalty. The paper never reports T's prediction error against the simulator on these OOD rollouts, nor does it show that eps_dyn improves, or at least does not degrade, success on OOD goals. The one framework ablation (Table 6) is on Door 30-90 only and shows Dyn+Joint (31.7%) barely outperforming Joint S&A (30.8%), so the dynamics model's contribution is already small even in-distribution. In addition, the paper's own analysis of the 70 task (8 of 14 failures opened to 90) illustrates that learned biases embodied by T can override goal guidance, exactly the failure mode expected if T extrapolates incorrectly. Without an OOD accuracy check, the claimed role of 'dynamics model-based dual guidance' in enabling robust adaptability is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DexHandDiff, an interaction-aware diffusion planner for dexterous manipulation. The method jointly diffuses over concatenated state-action trajectories and combines classifier-guided goal conditioning with a learned dynamics model, a dual-phase pre-contact/post-contact guidance mechanism, and LLM-generated guidance functions. The authors evaluate on five Adroit/Shadow Hand tasks with goal-adaptive variants, reporting a 59.2% average success rate across door-opening goal shifts (Table 2) and a 70.7% average across five tasks (Table 3), and argue that the approach reduces 'ghost states' by maintaining state-action consistency.","tokens_in":29897,"tokens_out":4087,"duration_ms":40097,"significance":"If the proposed mechanism is made formally sound, the paper would make a useful contribution to diffusion-based planning for contact-rich manipulation. The empirical evaluation is broad and the goal-adaptivity results, if reproducible and not dominated by tuned guidance hyperparameters, would be a meaningful advance over state-only and action-only diffusion baselines. The combination of dual-phase guidance, dynamics consistency, and LLM-generated guidance functions is a novel and potentially practical pipeline. The paper also provides a clear diagnosis of the ghost-state failure mode in state-based diffusion. However, the current presentation leaves two load-bearing technical points underspecified: the differentiability of the Heaviside-based guidance terms and the validity of the learned dynamics model outside the training distribution. These need to be resolved before the central claims can be assessed as fully supported.","major_comments":[{"comment":"The behaviour models h_penalty and h_finger are defined using Heaviside step functions. The classifier-guided formalism in Eqs. (7)-(8) and the product-of-experts derivation require the guidance gradient g = -Σ ∇_τ ε_i(τ,c) to be well-defined, which in turn requires the energies ε_i to be differentiable (or at least Lipschitz, as assumed in Section 3.3). The gradient of a Heaviside function is a Dirac measure, not a function; the statement that applying Eq. (11) makes ε_penalty 'a Dirac delta function that directly sets value when satisfying the constraints' does not resolve this, because a delta measure cannot be used as a gradient term in the Gaussian reverse-process update. The authors should either replace Eqs. (14) and (16) with smooth surrogate penalties that are actually used in implementation (e.g., sigmoid or soft-threshold terms), or provide a rigorous treatment of how step-function behaviour models are incorporated. This is load-bearing because the entire dual-guidance mechanism is presented as gradient-based classifier guidance.","section":"Section 5.1, Eqs. (14) and (16)"},{"comment":"The phase-transition threshold δ1 and the penalty thresholds δ2 and δ3 are never reported, despite being central to the dual-phase guidance; the text says a 'smooth transition mask' is applied, but Eq. (13) is a hard inequality based on δ1. In addition, Appendix D states that the guidance scale α is task-dependent, selected from {500, 1000, 2000}, but the per-task values used to produce Tables 2 and 3 are not given. Because these parameters directly control the strength of the guidance terms that produce the reported success rates, the results are not reproducible without these values, and it is unclear whether the reported averages reflect robust behaviour or manually tuned guidance gains.","section":"Section 5.1, Eq. (13), and Appendix D"},{"comment":"The learned dynamics model T(s,a), trained only on demonstration data, is used as the dynamics-consistency guidance ε_dyn at inference on out-of-distribution goals (e.g., door at 30°, door closing, pen left-hemisphere rotation, hammer half-drive). The paper never validates T's prediction error on these OOD settings, so it is not established that ε_dyn provides a correct penalty rather than a corrupting one. The paper's own ablation in Table 6 shows that Dyn+Joint (31.7%) barely improves over Joint S&A (30.8%), and Dyn-guide alone gives the same average as Diffuser (27.5%). Moreover, the analysis in Section 6.1 reports that 8 of 14 failures on the 70° task opened to 90°, which is consistent with learned dynamics overriding the goal guidance. The authors should report T's prediction error against the simulator on OOD rollouts, or provide an OOD ablation with and without ε_dyn, to support the claim that dynamics-based guidance enables robust adaptability.","section":"Section 5.1, Eq. (12), and Section 6.5, Table 6"}],"minor_comments":[{"comment":"Reference [30] is listed as 'Dexdiffuser: Interaction-aware diffusion planning for adaptive dexterous manipulation, arXiv:2411.18562, 2024', which is the same arXiv identifier as the present manuscript. Citing your own prior version as if it were independent external work is misleading; this reference should either be removed or clearly presented as a self-citation with an explicit description of the relationship.","section":"References, [30]"},{"comment":"The 'ghost state' metric in Table 4 is underspecified: the text does not define what 'predicted' and 'simulated' states are, how the L2 distance is computed and normalized beyond 'per dimension', or why Conditional DP is excluded. Without this detail, the quantitative claim that DexHandDiff 'nearly halves' the baseline's gap is hard to interpret.","section":"Section 6.3, Table 4"},{"comment":"The phrase 'over 3 tries for 10 random seeds' is ambiguous about the total number of evaluation episodes per task. Clarify whether this means 3 episodes per seed (30 episodes total) or some other configuration, and report the total episode count.","section":"Section 6.1"},{"comment":"There is a typographical formatting issue in the Decision Diffuser row for Open 50°, where '3.3 ± 4.7' appears as '3 .3 ± 4.7'.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'd want a referee to see this, but with a clear request for revision. The real contribution is the joint state-action diffusion planner that splits guidance into a pre-contact alignment phase and a post-contact goal phase, and the use of LLM-generated guidance functions to cut down on manual reward engineering. That combination is new for dexterous manipulation, and the experiments are set up to test exactly the right thing: trained on one goal, evaluated on shifted goals. Tables 2 and 3 are convincing as an empirical recipe, and Table 6 shows the interaction mechanism, not the dynamics model, is what drives the jump to 67.5% success.\n\nThe soft spots are real but mostly fixable. The formalization in Sec. 5.1 with Heaviside penalties (Eqs. 14 and 16) cannot be right as written: the gradient of a step function is a Dirac delta, which is not a usable classifier-guidance term. The appendix code, though, uses smooth L2 penalties. The paper should present the smooth functions actually used, or explicitly say the Heaviside is a conceptual simplification. Second, the dynamics model T in Eq. 12 carries more weight than the evidence supports. It is trained only on demonstrations, and the OOD goal rollouts require it to extrapolate. No error against the simulator is reported on those OOD states, and the 70-degree door failures (8 of 14 opened to 90) are consistent with the learned model overriding goal guidance. The Table 6 ablation shows Dyn+Joint at 31.7% vs. Joint S&A at 30.8%, so the marginal gain in-domain is tiny. The paper should validate T's prediction error on the adaptive-goal rollouts or soften the claim that dynamics-based dual guidance is what makes adaptation work.\n\nSmaller issues: delta1-delta3 are never given values; the guidance scale alpha is per-task tuned; no code or data is released; and Ref. [30] is the same arXiv ID under an old title (DexDiffuser), which should be corrected. The LLM guidance ablation honestly shows it underperforms human-crafted guidance, which is fine, but the abstract's phrasing overstates the automation.\n\nThe central claim, that interaction-aware dual-phase guidance improves goal-adaptive dexterous manipulation, is supported by the data. The paper is for researchers working on diffusion planning and contact-rich manipulation; it gives a concrete recipe with believable benchmark numbers and unusually detailed appendix material. I'd send it to review rather than desk reject. For a revision, I'd want the math aligned with the implementation, the dynamics model's OOD error reported, and code plus full hyperparameters released.","headline":"Empirically solid goal-adaptive dexterous manipulation planner; the guidance math and the dynamics-model evidence need revision before acceptance.","tokens_in":30469,"tokens_out":4094,"would_cite":true,"duration_ms":38892,"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":"DexHandDiff claims that diffusion planners for dexterous manipulation fail on contact-rich tasks because they generate objects as if they could move on their own, and that jointly denoising hand states, object states, and actions with…","keywords":["dexterous manipulation","diffusion planning","ghost states","classifier-guided diffusion","dynamics model","goal adaptation","LLM-generated guidance","contact-rich manipulation"],"falsifier":"Measure the learned dynamics model's one-step prediction error on guided trajectories for an out-of-distribution goal, such as 30-degree door opening, by comparing its predicted next state to the simulator's next state; if trajectories that still succeed contain large dynamics errors, then the dynamics guide is not doing the load-bearing work and ghost-state reduction should instead be attributed to the contact-phase mask.","tokens_in":29405,"feed_emoji":"🦾","tokens_out":5367,"duration_ms":51199,"temperature":0.7,"pith_summary":"DexHandDiff claims that diffusion planners for dexterous hands fail in contact-rich tasks because they generate either actions or object states alone. Generating object states directly lets objects appear to move without hand contact, which the paper calls ghost states; generating only actions cannot steer toward new goals without new data. The paper instead jointly denoises hand states, object states, and actions, using a learned dynamics model plus phase-aware guidance to keep state-action pairs physically consistent. On a door-opening model trained only on 90-degree demonstrations, DexHandDiff reports 59.2% average success across goal-shifted door tasks versus 29.5% for the next best baseline, and 70.7% average success over ten in-domain and goal-adaptive task variants. The significance is that one demonstration dataset can be repurposed to new goals and even reversed tasks, with LLM-written guidance functions replacing hand-designed rewards.","feed_headline":"DexHandDiff doubles dexterous-manipulation success over baselines","feed_subtitle":"Jointly modeling hand, object, and action kills ghost states and adapts one demo set to new goals.","key_machinery":"The load-bearing object is the joint state-action diffusion trajectory, where the denoised object is a concatenated sequence of hand states, object states, and actions. Multiple guidance objectives are combined as a product of experts, so the reverse-process gradient is the sum of individual guidance gradients. A separately learned dynamics model supplies a consistency energy, penalizing state-action pairs whose predicted next state disagrees with demonstrated dynamics. A contact-distance phase switch applies alignment guidance before contact and adds goal, dynamics, and physical-constraint guidance after contact, with the physical constraint implemented as a Heaviside bound on per-step object motion; in-hand tasks use goal, finger-motion, dynamics, and constraint guidance. A two-stage LLM process generates the differentiable guidance code from environment descriptions and task instructions.","core_discovery":"The central discovery is that the ghost-state failure is not a diffusion artifact but a controllability mismatch: object degrees of freedom are not directly actuated, so a diffusion model that independently denoises them invents object motion. DexHandDiff instead models the joint state-action trajectory and couples states and actions through a dynamics consistency energy in the sampling objective. A dual-phase schedule first guides the hand to the contact point, so no object motion is attempted before contact, and then guides the coupled hand-object system to the goal, with a step-size constraint bounding per-step object change. This lets a model trained only on 90-degree door opening reach 70.0% success on 30-degree opening and 40.0% on the left-hemisphere pen reorientation, against 16.7% and 23.3% for the strongest baselines on those tasks. The paper also shows LLM-generated guidance functions come within a few points of hand-crafted ones, reducing human tuning iterations from roughly twenty to about five.","pith_inferences":["The paper never validates the learned dynamics model's prediction error on out-of-distribution states, so the dynamics guidance is the most unexamined load-bearing component; an ensemble or a finetuning pass on augmented data would be a direct stress test.","Because each guidance function depends on explicit contact-point geometry, extending the method to novel objects likely requires automatic contact-point detection or letting the LLM infer contact structure directly from the environment description.","The in-domain accuracy penalty (90% versus 100% on the training goal) suggests users should expect a small performance tax when using goal-adaptive guidance, and scheduling guidance strength by task familiarity could recover some of it.","LLM-generated guidance lags hand-crafted guidance on door 30 degrees and pen half-side reorientation, so prompt and template refinement remains a clear lever for closing the remaining gap."],"forward_implications":["A single diffusion checkpoint can handle multiple goals of the same task if guidance is available, removing the need to recollect demonstrations for each new target configuration.","The state-action coupling plus dynamics guide reduces the gap between predicted and simulated hand-object states by roughly half on adapted tasks, supporting the claim that ghost states are suppressed.","LLM-generated guidance functions make new dexterous goals addressable through language instructions, reducing manual reward engineering from about twenty iterations to about five.","The architecture is light enough for real-time control: 3.96M parameters and 3.27 GFLOPS, running at 5-7 Hz natively and about 36 Hz with a faster solver and command interpolation.","There is a tradeoff between adaptation and in-domain accuracy: classifier-free baselines score 100% on the exact training goal (door 90 degrees) while DexHandDiff scores 90%, suggesting users trade some training-distribution performance for goal flexibility."],"supporting_citations":[{"why":"Supplies the state-only classifier-guided diffusion planner that exhibits ghost states and serves as a primary baseline.","marker":"[24]"},{"why":"Provides the classifier-free conditional diffusion baseline and the inverse-dynamics formulation that limits explicit goal conditioning.","marker":"[1]"},{"why":"Represents action-only diffusion and demonstrates the limitation that action-only planners cannot explicitly steer toward shifted goals.","marker":"[14]"},{"why":"Provides the Adroit dexterous hand environments and expert demonstrations for door, hammer, pen, and relocate tasks.","marker":"[41]"},{"why":"Establishes the text-to-reward paradigm with code execution feedback that DexHandDiff adapts for LLM-based guidance generation.","marker":"[54]"},{"why":"Shows that LLMs can generate complete reward functions from language, motivating the automated guidance-function pipeline.","marker":"[33]"},{"why":"Supplies the D4RL datasets containing the teleoperated demonstrations used to train the diffusion and dynamics models on Adroit tasks.","marker":"[18]"}],"fun_headline_variants":["DexHandDiff doubles success, erases ghost states","Interaction-aware diffusion doubles dexterous success","Ghost states conquered: DexHandDiff doubles success rate","LLM-guided diffusion doubles dexterous task success","Adaptive dexterous manipulation with dual-phase diffusion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The learned dynamics model is trained only on the original demonstrations, so it must stay accurate for the out-of-distribution states and actions that goal-guided sampling produces; if it is wrong there, the dynamics guidance will push generated trajectories away from physical reality instead of preserving it.","fun_headline_variants_meta":{"raw":{"variants":["DexHandDiff doubles success, erases ghost states","Interaction-aware diffusion doubles dexterous success","Ghost states conquered: DexHandDiff doubles success rate","LLM-guided diffusion doubles dexterous task success","Adaptive dexterous manipulation with dual-phase diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00068,"raw_usage":{"total_tokens":3099,"prompt_tokens":966,"completion_tokens":2133,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":2058}},"tokens_in":582,"tokens_out":2133,"duration_ms":14470,"temperature":1.0,"reasoning_tokens":2058,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:05:13.936384+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the learned dynamics model's one-step prediction error on guided trajectories for an out-of-distribution goal, such as 30-degree door opening, by comparing its predicted next state to the simulator's next state; if trajectories that still succeed contain large dynamics errors, then the dynamics guide is not doing the load-bearing work and ghost-state reduction should instead be attributed to the contact-phase mask.","supporting_citations":[{"cited_title":"Text2reward: Reward shaping with language models for re- inforcement learning","cited_arxiv_id":null,"evidence_quote":"Establishes the text-to-reward paradigm with code execution feedback that DexHandDiff adapts for LLM-based guidance generation."}],"review_version":1}