{"id":"2ffb2520-6433-41ee-868c-711b7dc283e6","arxiv_id":"2507.04790","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A two-step model-merging method transfers interaction knowledge from multiple motion datasets to a target domain, outperforming ensembling and domain adaptation at the same inference cost.","lead":"Robots that navigate around people often fail when moved to a new environment, because different datasets encode different ways that humans and robots interact. This paper proposes IMMP, a method that merges the trained parameters of several navigation models, each from a different dataset, to adapt to a new target domain at no extra inference cost.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Diverse-source merging vs. target-set checkpoint selection: missing single-source control leaves the central transfer claim underdetermined.","rationale":"Read in good faith, the paper is a plausible engineering contribution: Table 1 consistently favors IMMP over DA, ensembling, and prior merging baselines across three planners, and the method is not circular because evaluation is on held-out target validation. The reader's CONDITIONAL verdict is reasonable. My pass points to a different load-bearing gap. Eq. (1) and Algorithm 1 learn the merging weights on the target training set, so the optimization can always emphasize the single checkpoint most similar to the target. The paper's own Fig. 5 shows weights correlate with zero-shot source-target similarity, confirming that the method can act as a soft selector; it does not show that combining multiple sources is better than picking the best one. A single-source control is therefore the decisive experiment for the central claim of 'leveraging diverse motion datasets.' I also note the DIPP implementation deviation already flagged by the reader, and some copied table entries (e.g., DTPP and DIPP Ties Merging THOR rows) that should be rechecked, but the diversity control is the primary issue. If the control passes, the claim stands and the manuscript's contribution is genuine; if it fails, the headline should be weakened to target-loss-weighted checkpoint selection, not interaction merging. Verdict remains CONDITIONAL; the specific acceptance condition is new, so I mark agreement as partial.","tokens_in":20548,"tokens_out":9651,"duration_ms":114984,"concrete_test":"On SIT and THOR with GameTheoretic, keep Algorithm 1 fixed but: (a) build the checkpoint pool from each single source domain alone; (b) build a same-size pool from random intermediate checkpoints of one source; (c) report the zero-shot ADE of each source checkpoint and the entropy/number of checkpoints with non-negligible learned weight. If the best single-source pool or the random one-source pool comes within 5% of the reported IMMP ADE (0.3380 on SIT, 0.1165 on THOR), the diversity-driven merging claim is not established; the method then reduces to target-loss-based checkpoint selection, and the paper should be revised accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"IMMP's per-module merging weights w_{i,theta} are optimized on the target training set (Algorithm 1 lines 14-18; Eq. 1), so the merged model can collapse to a target-loss-weighted selection among source checkpoints. The paper never reports quantitative single-source target performance: Fig. 3 is qualitative, and no ablation removes source domains from the checkpoint pool. If the best single-source pool, or a pool of random checkpoints from one source, reaches the reported ADE under the same weight-learning procedure, then the claim that diverse interaction knowledge is merged—rather than one useful checkpoint being selected—is unsupported. The DIPP variant (suppl B: weights extracted 'at once from the unified module') further means module-level interaction transfer is not exercised on all three backbones. Thus the causal role of source diversity is the least secure link in the central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Interaction-Merged Motion Planning (IMMP), a two-stage method for transferring knowledge from multiple source trajectory datasets into a target-domain motion planner. In the first stage, checkpoints are collected from models trained on each source domain, including metric-optimal checkpoints and intermediate checkpoints along the optimization trajectory. In the second stage, per-module task vectors are weighted and summed into an initial parameter vector, with the weights optimized on the target training set. The method is evaluated on SIT and THOR target domains with three planning backbones (GameTheoretic, DTPP, DIPP), reporting improved ADE, collision rate, FDE, and miss rate relative to domain generalization, domain adaptation, ensembles, and existing model-merging baselines at inference cost ×1. A supplementary study adds a Zara2 target and ablations on checkpoint composition, merging granularity, checkpoint interval, and module grouping.","tokens_in":20616,"tokens_out":4410,"duration_ms":50936,"significance":"If the reported results are robust, IMMP would be a practically attractive way to exploit existing motion datasets: it avoids joint multi-dataset training, adds no inference cost, and shows strong target-domain numbers across several backbones. The paper ships a code release, reports results on multiple planners and targets, and includes a useful supplementary analysis of checkpoint composition and hyperparameter sensitivity. However, the central transfer claim is currently underdetermined by the experiments: the learned merging weights could in principle collapse to selecting one useful source checkpoint, and the paper does not provide the single-source control needed to rule this out. In addition, the absence of repeated-seed variability, the target-dependent selection of checkpoint intervals, and the inconsistent treatment of DIPP's module granularity all limit confidence in the general claim. These are fixable with additional experiments and clarifications, so the work is suitable for major revision.","major_comments":[{"comment":"The central claim that IMMP leverages diverse source datasets is underdetermined without a single-source control. Because the merging weights w_{i,θ} in Eq. (1) are optimized on the target training set, the merged model can down-weight all but one source checkpoint. The paper never reports the target performance of a pool built from each source domain alone (or from random checkpoints of a single source) under the same weight-learning procedure; Fig. 3 is qualitative and no ablation removes source domains from the pool. If the best single-source pool already achieves the reported ADE, the claim that diverse interaction knowledge is merged rather than that one useful checkpoint is selected is unsupported. Please add quantitative single-source ablations and an oracle/upper-bound comparison.","section":"Sec. 4.2, Table 1 and Algorithm 1"},{"comment":"The experimental comparison lacks repeated-seed variability, and the headline result appears to select the checkpoint interval C on the target domain. Table 1 reports single numbers with no error bars or seeds. Suppl. Table 6 sweeps C over 1, 2, 3, 10 and reports 'Ours (5)' as the selected value; there is no held-out protocol for choosing C, so the reported 0.3157 ADE may reflect target-set tuning. The table also labels this row as 'without finetuning' while the same number appears in Table 1 and Table 3 for IMMP+Finetune, which is internally inconsistent. Please report mean and standard deviation over at least three seeds and clarify the checkpoint-interval selection procedure.","section":"Table 1 and Suppl. Table 6"},{"comment":"The method is claimed to be architecture-independent and to merge distinct modules (θego, θsurr, θinter, θelse), but this is not exercised on all three backbones. Suppl. B states that for DIPP 'weights are extracted at once from the unified module combining the planner and predictor.' Therefore the DIPP row in Table 1 does not test the proposed interaction-level grouping, and the claim of general applicability across planning models is weaker than presented. Please either report DIPP results with the same module-level separation, or clearly scope the claim to backbones where such separation is feasible.","section":"Sec. 3.4 and Suppl. B"},{"comment":"The paper claims that IMMP 'mitigates domain imbalance and catastrophic forgetting,' but no experiment measures either phenomenon. The protocol collects source checkpoints once and does not sequentially update the model on source domains, so it avoids catastrophic forgetting by construction rather than mitigating it in a measurable sense; domain imbalance is also not directly evaluated. Please either add an evaluation that tracks source-domain performance after merging and fine-tuning, or soften the claims to avoid unsupported causal language.","section":"Abstract and Sec. 5"}],"minor_comments":[{"comment":"The IMMP row in Table 8 reports FDE 0.9580 and Miss Rate 0.6976 for the same SIT/GameTheoretic setting, whereas Table 1 reports FDE 0.7626 and Miss Rate 0.6446; please reconcile the discrepancy.","section":"Suppl. Table 8"},{"comment":"The row labeled 'Domain Generalization [17]' appears to describe the dataset-level approach of UniTraj, which is cited as [16] in the text; please align the citation labels.","section":"Table 1"},{"comment":"The statement that intermediate checkpoints 'tend to be more generalized across domain shifts' is an empirical claim that would benefit from a more direct analysis or citation, even though Table 3 provides supporting evidence.","section":"Sec. 3.3"},{"comment":"The caption states that w is 'the average of contributions of task vectors per epoch and per module,' but the exact definition is not given; please provide the formula or pseudo-code used to compute the displayed weights.","section":"Fig. 3 and Fig. 4"},{"comment":"The metric definitions mix notation for ground-truth positions and planned positions (x and g); please define them uniformly and make clear that the collision threshold is applied to corresponding time steps of the plan and the ground truth.","section":"Eq. (6)-(9)"}],"recommendation":"major_revision","confidential_remarks":"The missing single-source control is the main substantive risk: if it is not addressed, the paper's central claim could reduce to 'target set checkpoint selection among source checkpoints.' The strength of the empirical results, the code release, and the reasonable ablations make this worth an additional round rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: IMMP is a sensible, domain-aware extension of model merging to motion planning, and the headline results look credible. The real gap is that the paper never shows that the diversity of sources is load-bearing: since the merging weights are learned on the target set, the method could plausibly be doing clever checkpoint selection rather than genuine interaction transfer. That control is missing.\n\nWhat's new: instead of merging one checkpoint per dataset, they build a pool with metric-specific best checkpoints plus intermediate epochs, and they merge per module (ego, surround, interaction, decoder). The module grouping is justified by the hierarchy in planning networks, and Table 2 shows it beats model- or parameter-level merging. That is a useful contribution. They also ship code and compare against domain adaptation, ensembling, and standard merging methods, which is more thorough than most papers in this space.\n\nThe results: consistent gains over baselines on SIT and THOR across three backbones, with no inference overhead. The ablation on checkpoint composition (Table 3) supports the value of metric- and epoch-selection. The supplementary shows correlation between domain similarity and learned weights, which is suggestive but not quantitative.\n\nSoft spots, in rough order of importance:\n\n1. Missing single-source control. The stress-test gets this right. With weights fit on the target, a pool of checkpoints from the single best source—or a pool of random checkpoints from one source—might match the reported ADE. Without that ablation, the paper's central claim 'diverse interaction knowledge is merged' is unproven. Table 5 (metric-wise checkpoints from one source) is a step but not the needed control.\n\n2. Statistical rigor. No repeated seeds or error bars. The checkpoint interval C is swept and the best value used in the main table (Table 6). The authors note robustness, but the headline numbers are still selected post hoc.\n\n3. DIPP inconsistency. For DIPP, weights are extracted 'at once' from a unified module, so the interaction-level grouping isn't actually exercised on all three backbones. That weakens the generality claim.\n\n4. Catastrophic forgetting is claimed to be mitigated but never measured directly; no continual learning or domain-imbalance evaluation.\n\nNone of these are dealbreakers. The method is plausible, the code is available, and the baselines are fair. The missing single-source control is one focused experiment away, and it's important.\n\nWho should read it: anyone doing cross-dataset transfer for planning or prediction, and the model-merging crowd looking for a new application domain.\n\nRecommendation: send it to review. The core idea is worth a proper round, but it needs the single-source ablation and some error bars before I'd fully trust the 'diversity' story.","headline":"A well-motivated adaptation of model merging to motion planning with credible gains, but the 'diverse sources' claim needs a single-source ablation before it fully lands.","tokens_in":21205,"tokens_out":3261,"would_cite":true,"duration_ms":37469,"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":"Merging per-module task vectors transfers interaction knowledge across motion datasets, beating ensembles and adaptation at zero extra inference cost.","keywords":["motion planning","model merging","task vector","cross-domain adaptation","interaction modeling","autonomous navigation","trajectory datasets"],"falsifier":"Train the same planner on two source datasets, merge their task vectors with weights learned on half of a target dataset, and evaluate on the other half: if the merged model fails to beat a target-only baseline, or if re-running weight optimization on a different target split yields drastically different weights, the transfer claim is contradicted. A second counterexample would be a dataset pair where the best-per-metric checkpoints are mutually incompatible, causing catastrophic collision failure even though each alone performs well.","tokens_in":20282,"feed_emoji":"🤖","tokens_out":5560,"duration_ms":58376,"temperature":0.7,"pith_summary":"The paper claims that an autonomous planner can absorb knowledge from many motion datasets without retraining on them all: take checkpoints trained separately on each source dataset, compute their task vectors, and learn per-module scalar weights that combine those vectors into one model for the target domain. On small target datasets such as SIT and THOR, this merged model beats training only on the target, domain adaptation, ensembling, and naive model-merging baselines on displacement, collision, and miss-rate metrics, while keeping inference cost at x1. The method also provides a stronger initialization: fine-tuning the merged model further improves every metric. The practical promise is a cheap, data-efficient route to cross-domain planning that reuses existing checkpoints and avoids domain imbalance and catastrophic forgetting.","feed_headline":"Merging checkpoints, not datasets, adapts robot planners","feed_subtitle":"Weighted task-vector merges transfer interaction knowledge across datasets with zero added inference cost.","key_machinery":"The load-bearing object is the task vector, defined as the difference between a trained checkpoint and an initial model. IMMP collects these vectors from source-domain checkpoints, groups parameters into ego, surrounding, interaction, and remaining decoder parameters, and optimizes a separate linear weight per group and per task vector against the target-domain loss. A second mechanism is the checkpoint pool: intermediate checkpoints at fixed intervals plus a best checkpoint for each metric, so the pool contains multiple snapshots of interaction patterns rather than a single fine-tuned model per domain.","core_discovery":"The central discovery is that interaction-level information, not whole-model parameters, is what transfers across motion planning datasets. By partitioning a planner into ego-encoder, surrounding-encoder, interaction-encoder, and decoder groups, and merging each group's task vectors with its own learned scalar weights, IMMP preserves the feature hierarchy that connects agent behaviors to interactions. The checkpoint pool mixes intermediate checkpoints at fixed intervals with the best checkpoint per metric (ADE, FDE, collision rate, miss rate), so each source domain contributes multiple snapshots of interaction patterns. Empirical comparisons across GameTheoretic, DIPP, and DTPP backbones show that this interaction-level merge outperforms model-level merging, parameter-wise merging, ensembles, and domain adaptation on SIT and THOR targets.","pith_inferences":["Beyond the paper: if linear mode connectivity holds for planning networks, the same pipeline could adapt one planner to many target domains by relearning only the scalar weights, without retraining source planners.","The learned merging weights appear proportional to domain similarity, so one could predict the weights from zero-shot target performance and skip the weight optimization entirely.","Because metric-specific checkpoints capture different interaction styles (collision-heavy versus miss-averse), the pool construction could be tuned toward a target's safety requirements rather than averaged across metrics."],"forward_implications":["Merged models reach target-domain performance without accessing source data during adaptation, keeping adaptation cheap and avoiding catastrophic forgetting.","The same merged parameters can serve as an initialization; fine-tuning them on the target set improves all metrics further.","Merging granularity matters: merging per interaction module outperforms whole-model or parameter-wise merging, suggesting planners should be merged along their feature hierarchy.","The method is backbone-agnostic: it composes with GameTheoretic, DIPP, and DTPP architectures without changing inference cost."],"supporting_citations":[{"why":"Defines task vectors and the weighted-sum merging formula that IMMP builds on.","marker":"[23]"},{"why":"Supplies the ensemble baselines (WTA and averaging) that IMMP outperforms at the same inference cost.","marker":"[2]"},{"why":"Ties-Merging baseline representing parameter-conflict resolution approaches that fail in planning.","marker":"[74]"},{"why":"UniTraj domain adaptation baseline that merges multiple trajectory datasets into unified training.","marker":"[16]"},{"why":"Model soups averaging baseline used as a merging comparison.","marker":"[69]"},{"why":"GameTheoretic planning backbone, its objective, and its forecaster-planner checkpoint structure.","marker":"[33]"},{"why":"DIPP backbone; its unified planner-predictor module defines how task vectors are extracted.","marker":"[21]"},{"why":"DTPP backbone with tree-structured planner used as another evaluation model.","marker":"[22]"},{"why":"SIT target dataset used to measure transfer performance.","marker":"[4]"},{"why":"THOR target dataset used to measure transfer performance.","marker":"[53]"}],"fun_headline_variants":["Merge interactions, not models, for robust planning","Checkpoint merging beats ensembles and domain adaptation","Interaction-level task-vector merge transfers driving skills","Zero-cost merge adapts planners across motion datasets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that task vectors from independently trained planners can be linearly combined with learned weights, and that the module hierarchy (ego, surrounding, interaction, decoder) is a meaningful partition—a hierarchy the paper relaxes for DIPP by extracting weights at once from the unified planner-predictor module.","fun_headline_variants_meta":{"raw":{"variants":["Merge interactions, not models, for robust planning","Checkpoint merging beats ensembles and domain adaptation","Interaction-level task-vector merge transfers driving skills","Zero-cost merge adapts planners across motion datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000137,"raw_usage":{"total_tokens":1095,"prompt_tokens":834,"completion_tokens":261,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":450,"completion_tokens_details":{"reasoning_tokens":203}},"tokens_in":450,"tokens_out":261,"duration_ms":3376,"temperature":1.0,"reasoning_tokens":203,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:39:41.836655+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same planner on two source datasets, merge their task vectors with weights learned on half of a target dataset, and evaluate on the other half: if the merged model fails to beat a target-only baseline, or if re-running weight optimization on a different target split yields drastically different weights, the transfer claim is contradicted. A second counterexample would be a dataset pair where the best-per-metric checkpoints are mutually incompatible, causing catastrophic collision failure even though each alone performs well.","supporting_citations":[{"cited_title":"Ties-merging: Resolving interference 11 when merging models","cited_arxiv_id":null,"evidence_quote":"Ties-Merging baseline representing parameter-conflict resolution approaches that fail in planning."},{"cited_title":"Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing in- ference time","cited_arxiv_id":null,"evidence_quote":"Model soups averaging baseline used as a merging comparison."},{"cited_title":"A game-theoretic framework for joint forecasting and planning","cited_arxiv_id":null,"evidence_quote":"GameTheoretic planning backbone, its objective, and its forecaster-planner checkpoint structure."},{"cited_title":"Th ¨or: Human-robot navigation data collection and accurate motion trajectories dataset","cited_arxiv_id":null,"evidence_quote":"THOR target dataset used to measure transfer performance."}],"review_version":1}