acceptodds
Under review as a conference paper at ICLR 2027

The Kinematic Similarity Principle: A Shared Kinematic Distribution across Mobility Domains Enables Zero-Shot Trajectory Prediction

Abstract

Existing trajectory predictors are specialized to a single mobility domain, such as pedestrians, vehicles, vessels, or aircraft. Zero-shot transfer has so far been studied within one domain, where the target is a new city, dataset, or task, while the mobility domain remains fixed. When the mobility domain is held out, existing predictors do not consistently outperform simple constant-velocity extrapolation. This is because they learn the scale and context of the domain they are trained on rather than the structure of motion that transcends domains. We show that such a shared motion structure indeed exists in the form of a shared kinematic distribution of short-horizon motion. Once such motion is expressed at the scale of its domain, domains share a distribution of future displacement given past kinematics. We call this regularity the *kinematic similarity principle*. A predictor that targets this shared motion can therefore be trained on data from some domains and applied to a held-out domain. We propose KineTraj, a predictor that exploits this principle. It includes (i) a scale-invariant kinematic encoder that expresses windows at the characteristic scale of its domain; (ii) a domain-agnostic backbone that takes only the resulting normalized representation as input, employing no domain labels or absolute scales; and (iii) a so-called KineMix decoder that combines closed-form motion regimes through a learned gate and residual to rescale predictions to the scale of a domain. On a leave-one-domain-out benchmark encompassing six datasets spanning five mobility domains, KineTraj, with 0.59M parameters, reduces best-of-20 ADE by 18% and FDE by 22% relative to the strongest domain-specific predictor trained under the same protocol, and it matches or exceeds trajectory and time-series foundation models applied zero-shot. Code is available at https://anonymous.4open.science/r/kinetraj_code-0D48.

Primary area:
applications to robotics, autonomy, planning
Keywords:
trajectory prediction, motion forecasting, zero-shot transfer

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_MHFkOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1(−1)
Read review
Paper summary

This work addresses the kinematic similarity between mobility domains like pedestrians, vehicles, vessels and aircraft. This work proposes a predictor that exploit the kinematic similarity and achieves zero-shot transfer performance in each mobility domain.

Critical strengths

This work focuses on an interesting problem, which is zero-shot transfering to different mobility domains, by using the kinematic simlarity. The designed framework targets to adapt to the characteristic of different mobility domain and achieves with a relatively light weight model.

Critical weaknesses
  1. Normalization for baseline methods. As the scale in different mobility domain is a huge difference, are baselines also normalized by domain-specific scale? It would be helpful to show the advantage of the proposed method, if the normalization is applied in the preprocessing step for all baselines. Otherwise, will feel the main performance drop can come from scale difference.

  2. Table 4 provides the squared error for both zero-shot and in-domain. a) it would be helpful to provide same metrics for baseline methods, to show the effectiveness of the zero-shot transfer ability more clearly. b) Although the squared error could represent the prediction error, it is necessary to show the same for ADE/FDE value, for in-domain as well, as ADE/FDE is used as main evaluation metrics in this work.

  3. About comparison between Table 1 and 3. In Table 3, ETH/UCY and SDD are put in the same fold for evaluation. Intuitively, the performance in Table 3 can be worse, as the most similar data source (ETH/UCY, or SDD) is removed from training data. As shown in Table 3, for the proposed method, removing SDD makes performance on ETH/UCY worse, but removing ETH/UCY makes performance on SDD same or even better in FDE. And similar patterns happen in some baselines for both datasets, which shows adding similar human trajectory data for training does not improve, or even downgrade the prediction accuracy. It would be helpful to provide the analysis of the results.

  4. The motivation is not clear enough. In introduction, it says: a new sensing platform or agent class lacks a predictor, and collecting enough data takes time. There is no concrete example provides, and in the experiments, all the evaluated mobility domains like aircraft or vehicle already have good amounts of data. Also even the method is targeting sparse data domain, the proposed method still requires observed data for scale computing.

  5. Following in point 4, it would be helpful to include evaluation of data-efficiency, that in-domain models are trained with an increasing number of trajecotries in this domain, and compared with the zero-shot results. This could show that, up to how many trajectories provided in target domain, the proposed method under zero-shot transfer outperforms in-domain training.

Questions for authors

Same as in weaknesses.

6 comments

We thank the reviewer for the careful and specific comments. We address the five weaknesses in order below.

W1. All baselines are normalized by domain scale. We thank the reviewer for pointing this out. Appendix A.4 of the revision now describes how each baseline is normalized. EqMotion and SingularTrajectory operate on coordinates divided by a per-domain median step length. UniTraj uses a per-domain z-score of displacements. OmniTraj is run with its scale-normalization option. Chronos and Time-MoE standardize each window. The LSTM, Transformer, CVAE and GAN baselines use the same per-domain coordinate normalization as KineTraj. The gap in Table 1 therefore does not come from a difference in scale.

W2. We added an experiment that trains KineTraj and EqMotion in domain and compares each with its zero-shot version in ADE and FDE. An in-domain model is trained on the training split of the held-out dataset itself, under the same cap of 2,000 trajectories as each source domain. The results are given below as minADE₂₀ / minFDE₂₀, in metres and in nautical miles for CT-DMA. The zero-shot rows repeat Table 1 and, like the in-domain KineTraj row, are means of three seeds.

MethodETH/UCYSDDCT-DMAT-DriveGeoLifeAircraft
KineTraj zero-shot0.261 / 0.4730.349 / 0.5711.223 / 2.255368 / 62617.2 / 29.4813 / 1500
KineTraj in-domain0.228 / 0.3970.332 / 0.5550.797 / 1.394308 / 51815.0 / 24.6713 / 1334
EqMotion zero-shot0.326 / 0.6320.376 / 0.6641.875 / 3.880464 / 84719.5 / 33.01102 / 2190
EqMotion in-domain0.276 / 0.4970.382 / 0.6320.886 / 1.529312 / 53815.3 / 25.2818 / 1544

Zero-shot KineTraj is within 5–20% of its in-domain version in ADE on ETH/UCY, SDD, T-Drive, GeoLife and Aircraft, and 53% behind it on CT-DMA, the vessel fold. Without any training data from the target domain, it has a lower ADE and FDE than EqMotion trained on that domain on ETH/UCY, SDD and Aircraft, and EqMotion trained in domain is better on CT-DMA, T-Drive and GeoLife. With the same in-domain training data, KineTraj has a lower ADE and FDE than EqMotion on all six folds, by 1–17% in ADE. These results are in Appendix B.2.2 (Table 9) of the revision.

W3. Similar pedestrian data does help, and how much it helps depends on whether a method can also use the data of other families. Removing the other pedestrian set raises the error in 13 of the 16 comparisons between Table 1 and Table 4 of the revision. The four coordinate-only models (LSTM, Transformer, CVAE and GAN) rise by 13–96%. They transfer poorly across families, so the other pedestrian set is the only source they can use. KineTraj, EqMotion and SingularTrajectory are the three trained methods that transfer best in Table 1, and they change by only −2% to +5%, because the vessel, vehicle and aircraft data already give them what the other pedestrian set gives.

We checked this with a new nearest-neighbour analysis. For each evaluation window of a pedestrian set, we take the 50 most similar source training windows in KineTraj's normalized input space. The table gives the share of each source dataset, in percent, among these 50 windows, among the 50 nearest windows in raw coordinates, and among all source training windows.

Held-out setWindows countedOther pedestrian setCT-DMAT-DriveGeoLifeAircraftOther families in total
ETH/UCY50 nearest, normalized space12.723.630.916.816.087.3
ETH/UCY50 nearest, raw coordinates69.10.07.623.30.030.9
ETH/UCYall source windows4.516.634.034.010.895.5
SDD50 nearest, normalized space18.217.530.417.116.881.8
SDD50 nearest, raw coordinates64.60.016.119.40.035.4
SDDall source windows4.716.634.034.010.895.3

In the normalized space, 82–87% of the most similar windows come from the vessel, vehicle and aircraft sets. In raw coordinates the other pedestrian set would supply 65–69% of them, although it holds under 5% of the source windows.

The three cases that do not get worse are within about two seed standard deviations. For KineTraj on SDD the change is 0.000 in ADE and −0.011 in FDE, against seed standard deviations of 0.010 and 0.029. SingularTrajectory on ETH/UCY changes by −0.008 in ADE against 0.013, and UniTraj on SDD by −0.144 against 0.074. The results therefore give no evidence that similar data lowers the accuracy.

We have added this analysis to Appendix B.1 of the revision.

W4 & W5. We rewrote the motivation in the introduction around the cold-start period of a new domain, and we added two experiments that measure what this period requires and how long it lasts. When a port radar begins to track a new vessel class, or a drone fleet starts in its first city, a few tracks are available on the first day, but the hundreds of trajectories that training needs take weeks or months to collect. KineTraj is meant for this period.

In such a new domain, the three numbers that KineTraj needs, the characteristic scales of step length, speed and acceleration, can be computed from observed past windows. The table gives the change of the zero-shot minADE₂₀ when these scales are computed from n randomly drawn observed windows in place of all test windows, averaged over ten draws and three model seeds. A dash means that a draw contained only stationary windows, for which the step scale is zero.

Observed windows nETH/UCYSDDCT-DMAT-DriveGeoLifeAircraft
1+11.1%+14.1%+15.9%+34.0%––
2+6.0%+6.6%+7.6%+18.5%+30.3%+1.2%
5+0.7%−0.6%+2.1%+9.2%+11.0%+0.8%
10+1.3%+1.1%+1.0%+8.8%+11.3%+0.7%
20−0.3%+0.1%+0.6%+1.6%+1.3%+0.6%
50−0.2%−0.3%+0.3%+0.8%+2.3%+0.2%
100−0.4%−0.1%+0.2%+0.3%+1.0%+0.2%

With only 20 randomly drawn windows in place of all test windows, the zero-shot error changes by less than 2% on all six folds.

In addition, we added a data-efficiency experiment. For each held-out domain, KineTraj is trained in domain on the first N training trajectories of that domain, with N from 10 to the full split, and compared with the zero-shot model, which sees no trajectory of the domain. The results are given below as minADE₂₀ averaged over three seeds. The last row is the N at which in-domain training matches the zero-shot model, found by interpolation of the curve.

In-domain trajectories NETH/UCYSDDCT-DMAT-DriveGeoLifeAircraft
100.4570.5851.78056026.21185
300.4210.5621.47950624.11097
1000.2740.4721.16639619.6935
3000.2410.3611.02134417.5821
10000.2330.3460.86932916.7730
all0.2280.3320.79730815.0713
Zero-shot KineTraj (N = 0)0.2610.3491.22336817.2813
N at which in-domain matches zero-shot≈160≈800≈80≈180≈470≈330

With 10 labelled trajectories of the target domain, the in-domain model is 46–75% worse than the zero-shot one, and with 30 it is still 21–61% worse. In-domain training catches up only at about 80 to 800 trajectories, depending on the domain (last row of the table). The full curve is in Appendix B.4 (Table 12 and Figure 5) of the revision.

We hope this addresses all concerns and thank them again for the suggestions, which have strengthened the paper.

Thanks the authors for the detailed response. The additional results resolves W2, W3 and W5 well. For W1, the concern of the normalization still remains. The baselines are not evaluated with and without normalization. LSTM, Transformer, CVAE and GAN are reported to use the same normalization scale as the proposed method, but when on the in-range folds like ETH/UCY or SDD, their ADE is about 2 times that of constant velocity model, while on a very difference scale source, like Aircraft scenario, it jumps to about 10 times, which shows the normalization may not effectively remove the scale difference for these baselines. For W4, the practicality and real-world use case of the proposed method are still not convincingly motivated. For example, a new vessel class still belongs to the existing vessel domain, where existing vessel data would be the better and natural choice rather than cross-domain transfer from pedestrian data. I will keep my rating.

Reviewer_yUGDOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies zero-shot trajectory prediction across mobility domains. The authors attempt to transfer motion patterns learned from existing domains to different types of agents, such as pedestrians, vessels, vehicles, and aircraft. The authors propose the “kinematic similarity principle,” suggesting that motion in different domains can share some motion patterns after scale normalization. Based on this idea, the authors design KINETRAJ, which combines kinematic feature normalization, a shared network, and physical models to generate multiple candidate trajectories. The model does not require training or fine-tuning in the target domain, but it requires scale statistics from historical observations in the target domain during inference. The experimental results also show that the method achieves the best ADE on five of the six datasets and, under the best-of-20 metric, reduces ADE by approximately 18% and FDE by approximately 22% on average compared with the stronger of EqMotion or SingularTrajectory on each dataset. It has some research value.

Critical strengths
  1. The authors’ research question has some significance. The paper studies zero-shot trajectory prediction across different mobility domains, such as pedestrians, vessels, vehicles, and aircraft, examining transfer capability when motion scales differ substantially, which has some practical value.
  2. The model is also relatively lightweight, and the experimental results are good. KINETRAJ has only approximately 590,000 parameters, achieves the best ADE on five of the six datasets, and performs well on the aircraft dataset, whose motion scale is significantly beyond the training range.
Critical weaknesses
  1. The departure measure defined in the paper only measures differences between the conditional means of different domains, while similar conditional means do not imply similar full distributions and cannot establish that different domains have similar multimodal structures. The relationship between error and the departure measure is established through a squared-loss decomposition, but it does not explain whether this departure should be sufficiently small after scale normalization. Therefore, the theory still provides a limited explanation of the shared distribution emphasized in the paper and the multimodal prediction advantage under best-of-20.
  2. The paper reports ablations on scale settings and the backbone network, but it has not sufficiently distinguished the respective contributions of the physical reference trajectories, learned gating, and residual prediction. In particular, it lacks a direct comparison with the simplified method of “the same scale normalization + a constant-velocity reference trajectory + multiple candidate residuals.” In addition, the model’s output candidate set itself contains the constant-velocity trajectory, so performing no worse than the constant-velocity baseline under best-of-20 is guaranteed by the construction of the candidate set and cannot alone serve as evidence that the model has learned shared patterns across domains. More targeted comparisons with the same number of candidates are needed to verify the additional value of the full KineMix design.
Questions for authors
  1. The departure measure in the paper only measures differences in conditional means. Could the authors further explain whether the full conditional distributions of different domains are also similar after scale normalization? Is there theoretical or experimental evidence supporting this similarity and explaining its connection to best-of-20 multimodal prediction performance?
  2. With the same scale normalization and the same number of candidate trajectories, how much does the performance of the full KineMix differ from that of the simplified model using “a constant-velocity reference trajectory + multiple candidate residuals”? Could the authors provide this comparison to explain the additional contributions of multiple physical reference trajectories and learned gating?
4 comments

We thank the reviewer for the detailed comments. Our point-by-point responses follow.

W1 & Q1. After scale normalization, the full conditional distributions of the domains are close but not identical. To establish this, we first extend our theory with a distributional departure (Definition 3 of the revision), so that the principle is stated on full distributions, and with Proposition 3, which connects the best-of-20 error to this distributional departure. We then add an experiment. It shows that after normalization the full conditional distributions of a held-out domain and of the sources are close on all six folds.

A departure on full distributions. The departure of the submission, δd\delta_d (Definition 2), is the root-mean-square gap between two conditional means. Two conditionals with the same mean and different modes have δd=0\delta_d=0, so δd\delta_d alone cannot support a claim about a shared distribution. The revision adds the distributional departure of Definition 3,

δdW=(Ed[W1(Qd(⋅∣k′),Q(⋅∣k′))2])1/2,\delta_d^W=\Big(\mathbb{E}_d\big[W_1\big(Q_d(\cdot\mid k'),Q(\cdot\mid k')\big)^2\big]\Big)^{1/2},

in which Qd(⋅∣k′)Q_d(\cdot\mid k') and Q(⋅∣k′)Q(\cdot\mid k') are the distributions of the normalized future in the held-out domain and in the pooled sources given the same normalized past k′k', and W1W_1 is the Wasserstein-1 distance. This departure changes when the two conditionals differ in spread, in the number of their modes or in where the modes lie. It is never smaller than the mean departure, δd≤δdW\delta_d\le\delta_d^W (Lemma 1). Principle 1 is restated with both, as agreement in conditional mean up to δd\delta_d and in the whole conditional distribution up to δdW\delta_d^W.

The connection between the similarity of the full conditional distributions and best-of-20 multimodal prediction performance. Proposition 3 of the revision establishes this connection for best-of-K losses, including minADEK\mathrm{minADE}_K and minFDEK\mathrm{minFDE}_K. Take the set of K hypotheses that is optimal for the pooled sources and apply it to the held-out domain. Its error exceeds that of the set that is optimal for the held-out domain itself by at most 2δdW2\delta_d^W, in normalized units. This is how the theory accounts for the advantage under best-of-20. A set of hypotheses that covers the modes of the shared distribution also covers those of a domain close to it. The detailed proof is given in Appendix D of the revision.

Experiment. We also added an experiment that measures the distance between the full conditional distributions of each held-out domain and of the sources, without and with scale normalization (Appendix B.2.2 and Table 8 of the revision).

Setup. For each of up to 5,000 evaluation windows of the held-out domain, we take the 50 windows with the most similar past from two pools, the held-out domain's own training data and the pooled sources. Their futures give two samples of futures that follow a similar past, and we compute the exact W1W_1 distance between the two samples. Two finite samples have a non-zero W1W_1 even when they come from the same distribution, so we divide the distance by a floor. The floor is the same distance between two samples of 50 neighbours taken from two disjoint halves of the held-out domain's training data. A value of 1.0 therefore means that the sources are as close to the held-out domain as the domain is to itself. We run the measurement without normalization, with pasts and futures in metres, and with scale normalization, with pasts and futures divided by the characteristic scales of their own domain as in KineTraj.

Results. The table gives the distance over its floor for each fold.

SettingETH/UCYSDDCT-DMAT-DriveGeoLifeAircraft
Without scale normalization11.022.73.01.610.23.6
With scale normalization2.02.03.41.52.13.1

From these results we draw two key conclusions.

  1. Scale normalization makes the full conditional distributions comparable. Without it, the distance is 10–23 times its floor on ETH/UCY, SDD and GeoLife, where the futures of the source neighbours are 8–19 times more dispersed than those of the domain's own neighbours. With it, the distance is about twice the floor on these three folds. On CT-DMA, T-Drive and Aircraft the multiple is 1.6–3.6 without normalization and 1.5–3.4 with it.
  2. After scale normalization the distributions are close but not identical. All six folds lie between 1.5 and 3.4 times the floor.

W2 & Q2. We added a comprehensive ablation that separates the contributions of the physical reference trajectories, the learned gating and the residual prediction. All variants produce 20 candidates, and each changes one component of the full model. Its results are given below as minADE₂₀ averaged over three seeds, with the best variant in bold and the second best underlined.

ComponentVariantETH/UCYSDDCT-DMAT-DriveGeoLifeAircraft
Full modelKineTraj0.2610.349‾\underline{0.349}1.22336817.2813‾\underline{813}
Normalizationw/o normalization0.2750.3571.77147027.71130
Normalizationw/ one global scale for all domains0.2710.3551.57754127.1989
Decoderw/o KineMix (Gaussian decoder)0.3050.4112.16451132.42638
Decoderw/o physics anchors0.2780.3651.57640026.71704
Decoderw/ CV anchor only0.2510.3611.255‾\underline{1.255}37918.5840
Decoderw/o learned residual modes (replace them with Gaussian residual)0.2970.3701.38942418.8905
Decoderw/o learned gate (fixed on the CV anchor)0.260‾\underline{0.260}0.3411.258371‾\underline{371}17.4‾\underline{17.4}805

From this ablation we draw four key conclusions.

  1. Both the per-domain normalization and the KineMix decoder are necessary. KineTraj is 2–5% better than w/o normalization on ETH/UCY and SDD and 22–38% better on the other four folds, and a single scale shared by all domains does not close this gap. ETH/UCY and SDD lose little because each has the other, at the same raw scale, among its sources. KineTraj is also 14–15% better than w/o KineMix on ETH/UCY and SDD and 28–69% better on the other four folds.
  2. Inside KineMix, a CV anchor with learned residual modes brings the largest gain. w/ CV anchor only is 12–18% better than w/o KineMix on ETH/UCY and SDD and 26–68% better on the other four folds. Both parts count. It is 1–5% better than w/o physics anchors on SDD and T-Drive and 10–51% better on the other four folds, and KineTraj is 6–13% better than w/o learned residual modes on every fold.
  3. More physics anchors add domain-independent prior knowledge and lower the error further on five of the six folds. The anchors are closed-form kinematic models without learned parameters, so they do not depend on the source domains. With four anchors in place of the CV anchor alone, that is, full KineMix against "a constant-velocity reference trajectory + multiple candidate residuals", the error is 2–7% lower on SDD, CT-DMA, T-Drive, GeoLife and Aircraft and 4% higher on ETH/UCY.
  4. The learned gate does not lower the ADE measurably, but it makes KineTraj more robust. w/o learned gate is within 3% of KineTraj in minADE₂₀ on every fold. In minFDE₂₀, KineTraj is the best or the second best variant on every fold and never more than 0.6% behind the best one, whereas w/o learned gate is up to 4.2% behind (Table 2 of the revision).

Since the candidate set contains the CV trajectory, being no worse than CV is guaranteed by construction, and the revision now states this (Sections 5.2 and 5.3). w/o physics anchors has no CV trajectory among its 20 candidates, and it is still 32–55% better than CV (Table 1 of the paper) on five folds and 6% worse on the off-scale Aircraft fold.

Section 5.4 and Table 2 of the revision report this ablation in more detail, together with the FDE.

We hope this answers both points, and we are glad to provide further details.

I thank the authors for providing a detailed rebuttal. The authors have addressed my main concerns, so I maintain my rating of 3.

We are glad that our response addressed your main concerns. Thank you for your constructive feedback and for your positive rating.

Reviewer_NmNsOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper addresses the problem of zero-shot trajectory prediction using models trained on other mobility domains, without requiring labeled trajectory data from the target domain. Building on the proposed Kinematic Similarity Principle, the paper introduces KineTraj, which consists of a scale-invariant encoder, a domain-agnostic backbone, and a KineMix decoder that combines four physics-based motion anchors with learned residuals. Through a leave-one-domain-out benchmark, the proposed method achieves competitive performance, demonstrating the potential for effective zero-shot generalization across heterogeneous mobility domains.

Critical strengths

• Unlike prior work, the paper addresses zero-shot generalization across different types of moving agents, which is a meaningful research direction. The proposed approach could be particularly useful in domains such as aircraft, where trajectory data may be difficult to obtain.

• The model achieves the lowest best-of-20 ADE on five out of six datasets. On the aircraft fold, it achieves a 26% improvement over EqMotion, demonstrating effective generalization even when the absolute motion scale differs substantially from those of the training domains.

• Based on the proposed Kinematic Similarity Principle, the model combines a scale-invariant encoder, physics-based anchors, and learned residuals in an intuitive manner. It also achieves strong performance with a relatively small model size (0.59M parameters), and its performance remains stable across different backbone architectures.

Critical weaknesses

• Proposition 1 provides a guarantee concerning the difference between the conditional means of the source and target domains, whereas the experiments primarily evaluate performance using multimodal best-of-20 ADE/FDE metrics. This theoretical result alone is insufficient to establish the similarity of the kinematic distributions across domains.

• It is difficult to determine whether the performance improvements primarily stem from scale normalization or the architecture of the KineMix decoder. Additional ablation studies would help better isolate the contributions of individual components.

• The paper does not provide qualitative visualizations comparing predicted trajectories with ground truth trajectories. Furthermore, since the best-of-20 metric considers only the prediction closest to the ground truth among multiple generated hypotheses, it needs to examine whether the improvements could also be attributed to broader hypothesis coverage.

Questions for authors
  1. Under the same experimental setting, how much additional performance improvement does the KineMix decoder provide compared with a model that uses only CV anchor and normalized residuals, rather than four physics-based anchors?

  2. Could the authors provide qualitative examples of trajectory predictions? In particular, I would like to see visualizations of the predicted trajectory distributions to better understand whether the improvements in best-of-20 performance arise from meaningful multimodal predictions or simply from broader hypothesis coverage.

  3. Could the authors explain more concretely how the departure (δ_d), defined based on the conditional means, supports the similarity of the full multimodal conditional distributions?

3 comments

We thank the reviewer for the constructive review. The three weaknesses and three questions come down to three concerns, and we address them one by one.

W1 & Q3. To address the gap between Proposition 1 and the best-of-20 metric of the experiments, we extended the theory. Alongside the mean departure of Definition 2, we add a distributional departure (Definition 3 of the revision), which expresses our core idea of a shared distribution directly. We also added an experiment that supports this distribution-level claim on all six folds. Both are described below.

Extended theory. Definition 2 compares a held-out domain with the shared distribution only through their conditional means. Definition 3 compares the two conditional distributions as a whole:

δdW=(Ed[W1(Qd(⋅∣k′),Q(⋅∣k′))2])1/2,\delta_d^W=\Big(\mathbb{E}_d\big[W_1\big(Q_d(\cdot\mid k'),Q(\cdot\mid k')\big)^2\big]\Big)^{1/2},

where W1W_1 is the Wasserstein-1 distance between the distributions of the normalized future in the held-out domain and in the pooled sources, given the same normalized past k′k'. δdW\delta_d^W is therefore sensitive to the spread of the two conditionals and to the number and location of their modes. With it, Principle 1 now states that domains agree with the shared distribution in conditional mean up to δd\delta_d, and as whole conditional distributions up to δdW\delta_d^W.

We also add Proposition 3, which links the principle to the metrics of the experiments. Under any best-of-K loss, including minADEK\mathrm{minADE}_K and minFDEK\mathrm{minFDE}_K, the hypothesis set that is optimal for the sources, applied to the held-out domain, has an error at most 2δdW2\delta_d^W above that of the set that is optimal for the held-out domain itself, in normalized units. A domain close to the shared distribution therefore loses little in these metrics under zero-shot prediction, and this is what the experiments test. The full proof is in Appendix D of the revision.

Additional experiment. The similarity of the full conditional distributions across mobility domains is also measured directly (see our reply to W1 & Q1 of Reviewer yUGD; more detail in Appendix B.2.2 and Table 8 of the revision). We estimate δdW\delta_d^W on every fold and report it as a multiple of a floor. The floor is the value obtained when the held-out domain is compared with itself, using two halves of its own data. A multiple of 1 would mean no departure. Without normalization the multiple is 10–23 on ETH/UCY, SDD and GeoLife, and after normalization it is 1.5–3.4 on all six folds, which supports this similarity. The revision states this in Section 5.3.

W2 & Q1. We added an ablation that changes one component at a time, either the normalization or one part of the KineMix decoder. The results are as follows (minADE₂₀, mean of three seeds; best in bold, second best underlined):

ComponentVariantETH/UCYSDDCT-DMAT-DriveGeoLifeAircraft
Full modelKineTraj0.2610.349‾\underline{0.349}1.22336817.2813‾\underline{813}
Normalizationw/o normalization0.2750.3571.77147027.71130
Normalizationw/ one global scale for all domains0.2710.3551.57754127.1989
Decoderw/o KineMix (Gaussian decoder)0.3050.4112.16451132.42638
Decoderw/o physics anchors0.2780.3651.57640026.71704
Decoderw/ CV anchor only0.2510.3611.255‾\underline{1.255}37918.5840
Decoderw/o learned residual modes (replace them with Gaussian residual)0.2970.3701.38942418.8905
Decoderw/o learned gate (fixed on the CV anchor)0.260‾\underline{0.260}0.3411.258371‾\underline{371}17.4‾\underline{17.4}805

From this ablation study, we can get two conclusions.

  1. Per-domain normalization and the KineMix decoder are both necessary. Compared with w/o normalization, per-domain normalization lowers the error by 2–5% on ETH/UCY and SDD, where each pedestrian set has the other, at the same raw scale, among its sources, and by 22–38% on the other four folds; w/ one global scale for all domains recovers little of this, so the scale has to be the domain's own. Compared with w/o KineMix, the decoder lowers the error by 14–15% on ETH/UCY and SDD and by 28–69% on the other four.
  2. Every component of KineMix is needed, for accuracy or for robustness. The physics anchors lower the error by 4–52% and the learned residual modes by 6–13%, on every fold. Four anchors in place of the CV anchor alone lower it by 2–7% on five folds and raise it by 4% on ETH/UCY. The learned gate changes the ADE by less than 3%, and its benefit is robustness. With it, KineTraj is the best or the second best variant in FDE on every fold, at most 0.6% behind the best, whereas w/o learned gate falls up to 4.2% behind.

The revision states this in more detail in Section 5.4 (Table 2), which also reports FDE.

W3 & Q2. We added the visualization as Figure 4 in Appendix B.3 of the revision. For two example windows per fold, it shows the observed track, the ground truth and the 20 hypotheses of EqMotion, CVM-S and KineTraj side by side. KineTraj's gain comes partly from spread, but not only from it. Compared with EqMotion, the strongest trained baseline, KineTraj reaches a lower error with a smaller spread on four of the six folds. In the examples, EqMotion's hypotheses form a tight bundle along the observed heading plus one or two that run far away, which is where its larger spread comes from. CVM-S draws straight lines of equal length around the heading. KineTraj's hypotheses fan out around the manoeuvre the agent makes.

We also added a quantitative comparison of the error and the spread of KineTraj and the baselines. Each cell gives minADE₂₀ and, in parentheses, the spread of the 20 hypotheses. The spread is the mean pairwise distance between their final points [1, 2], in units of the fold's constant-velocity FDE. The CV and CVM-S rows are from Table 1. CVM-S (same spread) is an average over 20 random draws. EqMotion and KineTraj are means of three seeds, as in Table 1.

MethodETH/UCYSDDCT-DMAT-DriveGeoLifeAircraft
CV (one hypothesis)0.569 (0)0.808 (0)2.872 (0)854 (0)39.2 (0)1612 (0)
CVM-S0.399 (1.58)0.590 (1.22)1.411 (1.57)697 (0.59)33.2 (1.15)1265 (4.52)
CVM-S (same spread)0.318 (2.64)0.426 (2.63)1.482 (2.34)453 (1.75)22.2 (5.46)1015 (6.27)
EqMotion0.326 (5.24)0.376 (3.33)1.875 (4.48)464 (1.30)19.5 (3.70)1102 (9.39)
KineTraj0.261 (2.64)0.349 (2.63)1.223 (2.34)368 (1.75)17.2 (5.46)813 (6.27)

CV is a single constant-velocity line. CVM-S [3] is 20 constant-velocity lines with random heading noise of a fixed size. CVM-S (same spread) is CVM-S with its noise matched to KineTraj. It keeps the constant-velocity line and adds 19 random lines. Each random line ends at the constant-velocity end point plus a Gaussian offset, and the offsets are scaled on every window so that the 20 lines have the same spread as KineTraj's 20 hypotheses on that window.

From this, we can get following two conclusions.

  1. The baselines and KineTraj all benefit from spread. Spreading 20 constant-velocity lines (CVM-S) lowers the error of the single line (CV) by 15–51%. Giving CVM-S the spread of KineTraj lowers its error by a further 20–35% on five folds, but not on CT-DMA, where it is 5% higher. Part of KineTraj's gain therefore comes from spread.
  2. KineTraj does better with less spread, so its gain does not come from spread alone. Its error is lower than EqMotion's on all six folds, with a smaller spread on four (ETH/UCY, SDD, CT-DMA, Aircraft). On T-Drive and GeoLife its spread is larger. With the same spread as KineTraj on every window, CVM-S (same spread) is still 21–29% worse on all six folds, and widening CVM-S by its heading noise alone, up to 90°, never brings it closer than 12%. This remaining gain comes from where KineTraj places its hypotheses.

The revision reports this comparison in Appendix B.3 (Table 11).

We hope these additions address the concerns and would be glad to clarify anything further.

[1] Y. Yuan and K. Kitani. DLow: Diversifying Latent Flows for Diverse Human Motion Prediction. ECCV, 2020.

[2] J. Chen, J. Cao, D. Lin, K. Kitani, and J. Pang. MGF: Mixed Gaussian Flow for Diverse Trajectory Prediction. NeurIPS, 2024.

[3] C. Schöller, V. Aravantinos, F. Lay, and A. Knoll. What the Constant Velocity Model Can Teach Us About Pedestrian Motion Prediction. IEEE Robotics and Automation Letters, 5(2):1696–1703, 2020.

Thank you again for your careful review. As the discussion period is about to close, we would like to check whether our response and the revised paper have addressed your concerns.

If any point remains open, we would be glad to answer it in the time that is left. If your concerns are resolved, we would be grateful if you could consider updating your rating.

Comments on the paper

2 comments
Area_Chair_6LPiOpenReview ↗

please engage

Dear reviewers,

Please join the discussion with the authors.

Best, AC

Dear Area Chair and Reviewers,

We sincerely thank the Area Chair and all reviewers for the constructive comments and for the continued discussion. This feedback has substantially improved the paper. The revision extends the theory to the best-of-20 metric, adds a component ablation and an in-domain comparison, strengthens the coordinate baselines, and states the motivation more clearly.

We believe that the revision addresses the concerns of all three reviewers. Below is a short summary of where the discussion stands.

Reviewer yUGD has confirmed that our response addressed the main concerns.

We have not yet heard back from Reviewer NmNs. However, two of the three concerns of this reviewer, the gap between the theory and the best-of-20 metric and the request for an ablation, coincide with those of Reviewer yUGD, and we answered them with the same additions. The third asked for a visualization of the predictions, which is now Figure 4 in Appendix B.3. We therefore believe that these concerns are addressed as well.

Reviewer MHFk raised five points and has confirmed that three of them are resolved. For the remaining two, we provided a second response. We improved the four coordinate baselines, tested them with and without normalization, and updated Tables 1 and 4. We also made the motivation in the Introduction clearer.

With these improvements, we believe that the paper now meets the bar for acceptance, and we hope that you will consider supporting it.

Thank you for your time and for reviewing and handling our submission.

Best regards,

The Authors of Submission 15407

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.