acceptodds
Under review as a conference paper at ICLR 2027

CooperTrace: Cooperative Multi-Modal End-to-End Motion Prediction

Abstract

Modern autonomous driving stacks transition from modular pipelines to end-to- end learning frameworks. Equipped with diverse sensing modalities, autonomous vehicles are increasingly robust to sensor degradation and long-tail events. In the meantime, leveraging Vehicle-to-everything (V2X) communication, recent cooperative autonomy research also expands from perception to downstream tasks, demonstrating the benefits of situational awareness beyond individual sensing range. However, it is unclear how multi-modal sensing would interact with the end-to-end learning paradigm in the V2X setting. In this paper, we present CooperTrace, a cooperative, multi-modal, and end-to-end motion prediction framework to explore the feasibility and efficacy of this three-fold combination. CooperTrace fuses LiDAR and camera at every agent, attends to the features of connected agents, and trains the forecaster on tracking results from previous frames. Extensive evaluations on both simulated and real-world cooperative datasets show that CooperTrace achieves up to 68% lower motion prediction error, 134% higher MOTA, and 18% higher 3D detection AP than the compared state-of-the-art methods. We further stress-test its cooperative gains under emulated sensor degradation. Cooper Trace’s multi-modal multi-agent fusion maintains stable performance across these settings and degrades gracefully when one modality becomes unavailable, demonstrating its potential for harsh real-world deployment conditions

Primary area:
applications to robotics, autonomy, planning
Keywords:
Cooperative Perception, LiDAR-Camera Fusion, Motion Forecasting, End-to-End Autonomous Driving, V2X
Communication

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_PPNLOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposed CooperTrace, which have a specific position in the intersection of cooperative, multi-modal, and end-to-end motion prediction. Extensive evaluations on several cooperative datasets show that CooperTrace's good performance. The paper also provides stress-test performance on cooperative gains under emulated sensor degradation, which is meaningful for real-world applications.

Critical strengths
  • This paper provides valuable empirical analysis on the tracking-to-forecasting distribution gap. It proposes an explicit tracking history, and demonstrate the different effects of subcomponents. The findings provide transferable design insights for other motion forecasting systems, which may benefit even without adopting CooperTrace model.
  • The evaluation is relatively comprehensive with several real-world and simulated datasets. And authors use more comprehensive metrics like coverage, EPA, etc., beyond naive ADE, especially the same-population analysis helps demonstrate model capability in different-difficulty cases.
Critical weaknesses

This paper positions end-to-end framework as one of the core contributions, but Table 6 shows that joint fine-tuning only improves minADE from 1.365 to 1.338 m. And the main OPV2V result in Table 2 appears to correspond to the staged LiDAR-camera configuration in Table 6, rather than the jointly fine-tuned model. Why is the jointly fine-tuned multi-modal model not used for the main results? Overall, although the authors demonstrate that joint backpropagation is feasible, but current experiments cannot sufficiently show that such end-to-end backpropagation is the key source of performance gains.

Questions for authors
  • Following the mentioned critical weakness, can authors provide the end-to-end forecasting ablations (particularly the contribution of joint fine-tuning) on the other real-world datasets? Current experiment results may be not distinguishing enough to show the separate contributions.
  • Given the large improvement from explicit tracking histories in Table 7, how well would a history-only forecaster perform? In particular, it would be useful to know how much additional benefit the track queries and BEV features provide once the motion history is available.
  • How are the low, moderate, and high severity levels defined for the "Camera missing" and "LiDAR missing" corruptions? The implementation and corruption parameters for each level are not clearly specified.
2 comments

Thank you for the careful review, for recognizing our analysis of the tracking-to-forecasting gap, and for the focused questions on the end-to-end claim. We address each point below. Changes in the revised PDF are in blue.

1. The contribution of joint fine-tuning, and the model reported in Table 2 (W1, Q1)

Please see point 2 of the general response. For the joint fine-tuning ablation on real-world data, the revised jointly fine-tuned model on V2X-Seq improves over staged training in minADE@5s (1.036 → 1.019 m), coverage (0.780 → 0.803) and EPA (0.471 → 0.480). Table 2 now reports the jointly fine-tuned model.

In the revised PDF, Sec. 1 is reworded, Table 2 reports the jointly fine-tuned model, Table 8 keeps the cached-history row, and Appendix I reports the staged and jointly fine-tuned numbers.

2. A history-only forecaster, and what the track query and BEV features add (Q2)

Please see point 1 of the general response. In short, a history-only forecaster already performs well (minADE@5s of 1.295 m in the cooperative setting), BEV features contribute about 0.11 m, and the track query raises EPA from 0.422 to 0.450 once it is fused through a learned gate.

In the revised PDF, Table 9 holds the cooperative rows and appendix Table 22 the single-agent ones, Sec. 3.4 describes the gated fusion, and Sec. 4.5 explains the result.

3. Severity levels of the corruptions (Q3)

At low / moderate / high severity, "camera missing" blanks each camera view (the whole image set to zero) independently with probability 0.3 / 0.7 / 1.0, so at high severity no camera image remains, and "LiDAR missing" keeps a random 10% / 3% / 1% of the points. The new appendix Table 21 lists every corruption.

Corruption│LowModerateHigh
Camera missing│each view blanked w.p. 0.300.701.00 (all blank)
LiDAR missing│keep 10% of points3%1%
Beam reduction│keep every 2nd beamevery 8thevery 32nd
Point dropout│drop 10% of points50%90%
Fog (extinction)│0.005 m⁻¹ (≈780 m visibility)0.015 m⁻¹ (≈260 m)0.060 m⁻¹ (≈65 m)
Sector occlusion│60° wedge90°120°

In the revised PDF, Appendix D lists every corruption's parameters per severity.

Summary. We revised joint fine-tuning so that the forecaster trains on live tracker outputs, repeated the ablation on real-world V2X-Seq and report the jointly fine-tuned model in Table 2, added the history-only forecaster and a gated query fusion, and specified the severity levels. We would be glad to answer any follow-up questions.

Thank you for the detailed response.

For W1 and Q1, the additional comparisons clarify the contribution of end-to-end joint training, although the improvements remain marginal.

For Q2, the new experiments convincingly demonstrate the importance of explicit history, and I appreciate the authors' honest reporting, although the results also somewhat weaken the contribution of track queries.

Overall, my assessment remains somewhat critical. Nevertheless, I still find the authors' exploration and transparent empirical findings meaningful, and I will maintain my original rating.

Reviewer_kR6nOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents COOPERTRACE, a cooperative (V2X), multi-modal, end-to-end motion forecasting framework. The authors identify that existing cooperative autonomous driving models suffer from either single-modality limitations, non-differentiable modular tracking, or exposure bias from being trained on idealized ground-truth histories. To address these, COOPERTRACE introduces a differentiable query-propagation tracker combined with a motion head that consumes both implicit track queries and explicit tracker-generated coordinate histories. The authors also contribute a rigorous audit of prior baselines, exposing non-deployment evaluation shortcuts (e.g., oracle pre-filtering) and establish a unified evaluation protocol. The resulting system demonstrates substantial empirical improvements in trajectory prediction error, coverage, and robustness to sensor degradation.

Critical strengths
  • Protocol audit and deployment-realistic evaluation. The most valuable contribution is the audit of prior cooperative forecasting baselines. The authors show that CMP filters detections with ground truth before tracking and re-keys tracks to ground-truth identities (Table 18). They also show that UniV2X’s headline accuracy score masks a ~2.86 m absolute error. Re-evaluating all methods on raw tracker output, with coverage and EPA reported alongside error, gives the V2X community a more honest benchmark. I expect this to matter beyond this paper.
  • Modality independence is well validated. Using LiDAR only as training-time depth supervision keeps the camera branch independent at inference. The degradation experiments support this design: under moderate beam reduction, the single-agent fusion model forecasts 61.7% of eligible vehicles versus 16.0% for LiDAR-only (Table 4). The authors also report limits honestly: the redundancies do not compose when both sensors degrade, and the camera improves coverage rather than per-agent accuracy.
Critical weaknesses
  • Limited methodological novelty, framed as conceptual contribution. This is a strong systems-engineering paper rather than a new method. The two headline ideas are established in the forecasting literature. Feeding explicit kinematic histories alongside learned agent features is standard (e.g., VectorNet, MTR). Training the forecaster on upstream tracker output to avoid exposure bias is the core idea of tracking-in-the-loop prediction (PnPNet). PnPNet is cited, but not in connection with the “closed-loop learning” contribution. The paper’s own ablations reinforce this reading: joint fine-tuning of the full graph improves minADE@5s by only 0.027 m (Table 6), and the conclusion attributes the main gains to history alignment rather than end-to-end backpropagation. I would view the paper more favorably if it were positioned as an empirical correction of recent query-centric V2X baselines, not as an algorithmic advance.
Questions for authors
  • Positioning. Given PnPNet and the broader use of explicit histories, is the contribution meant as an algorithmic novelty, or as an empirical finding that recent query-centric cooperative systems (UniAD-style, UniV2X) lose accuracy by discarding explicit coordinates and realistic tracker noise? The answer affects how I weigh the “first end-to-end” claim.
  • Inconsistent degradation claim. The introduction says a “LiDAR-only cooperative agent” forecasts 16.0% under severe LiDAR degradation. That figure appears to be the single-agent, moderate-severity result in Table 4. Table 16 reports 0.327 for cooperative LiDAR-only under beam reduction, and the corresponding fusion cell was not run. Could the authors correct this and report the cooperative fusion result?
  • Latency and synchronization. Partner BEV maps are warped into the ego frame and fused by attention. Does the evaluation assume zero transmission latency and synchronized timestamps? How does fusion perform under realistic delays (e.g., 50–200 ms) or asynchronous partner messages?
2 comments

Thank you for the careful review, for recognizing the protocol audit and the modality-independent design, and for the positioning question. We address each point below. Changes in the revised PDF are in blue.

1. Contribution and positioning (W1, Q1)

Please see point 2 of the general response. In short, we now present the contribution as a system together with the empirical finding you describe, and we cite PnPNet, VectorNet and MTR next to it.

In the revised PDF, Sec. 1 states the combination and then the finding, Sec. 2 cites VectorNet, MTR, PnPNet and V2XPnP, Table 2 reports the jointly fine-tuned model, and Appendix I reports the staged and jointly fine-tuned numbers.

2. The 16.0% figure and the missing cooperative fusion cell (Q2)

Thank you for catching this. The introduction now reports the cooperative moderate-severity result, 32.7% of eligible vehicles for cooperative LiDAR-only under beam reduction against 63.1% for cooperative fusion (Table 4), and the single-agent numbers (16.0% and 61.7%) stay in the appendix (Table 17). We also rechecked every number in the paper against its table.

We ran the missing cooperative fusion cell as well. Under beam reduction, cooperative LiDAR + camera forecasts 63.1% of vehicles (EPA 0.238), against 32.7% for cooperative LiDAR-only (EPA 0.126). The revised appendix Table 18 is below (moderate severity, every agent corrupted).

│Coop. LiDAR│Coop. LiDAR + camera
Corruption│minADE@5s ↓Cov. ↑EPA ↑│minADE@5s ↓Cov. ↑EPA ↑
None (clean)│1.3650.9110.417│1.4840.8880.392
Fog attenuation│1.3470.7810.378│1.4750.8090.367
Beam reduction│1.5340.3270.126│1.7760.6310.238
Point dropout│1.3860.8640.402│1.5100.8600.376
Gaussian noise│1.3670.8970.400│1.5030.8880.389
Snow│1.4530.6640.286│1.5290.7480.321
Sector occlusion│1.5060.8140.331│1.6070.8510.350
LiDAR missing│1.6330.1870.065│1.8200.5790.211
Motion blur (camera)│1.3650.9110.417│1.4910.8540.376
Fog + motion blur│1.3470.7810.378│1.4850.7330.345

In the revised PDF, Sec. 1 is corrected, Tables 4 and 5 now report the cooperative models, appendix Table 18 is complete, appendix Table 19 extends it to low / moderate / high severity, and Sec. 4.3 adds a qualitative example (Figure 4).

3. Latency, synchronization and communication cost (Q3)

Thank you for raising this. The main tables assume zero latency and synchronized partners, so we added a separate latency study in the new Sec. 4.4. We test partner delays of 100 and 200 ms and asynchronous partners. On cooperative OPV2V with LiDAR and map, a model trained without delay degrades at 200 ms and when each partner is delayed independently by a random 0–200 ms (Async). With delay-augmented training (random 0–2 frame partner delays), minADE@5s and EPA remain almost unchanged in both cases. The revised Table 7 with the asynchronous column is below.

│minADE@5s ↓│EPA ↑
Training│0 ms100 ms200 msAsync│0 ms100 ms200 msAsync
Without delay│1.4171.5101.6341.644│0.3930.3570.1060.220
Delay-augmented│1.4341.4511.4561.444│0.3900.3850.3780.385

We also tried feeding each partner's delay to the fusion as a learned embedding (delay-aware fusion). It leaves detection unchanged, so exposure to delay during training is what matters.

│mAP_BEV ↑
│0 ms100 ms200 msAsync
Delay-augmented│0.9670.9630.9550.963
+ delay-aware fusion│0.9670.9630.9550.963

For bandwidth, top-k sparsification and int8 quantization in training cut the exchange from 18 MB to 13.5 KB per partner per frame for 1.5 points of detection mAP. Forecasting with the unchanged tracker and motion head barely moves (minADE@5s 1.405 → 1.410 m, EPA 0.406 → 0.395). The new Table 6 is below.

ExchangePayload / partner / frame│mAP_BEV ↑mAP_3D ↑
Dense fp1618 MB│0.9690.903
Top-k 0.5% + int850 KB│0.9550.899
Top-k 0.15% + int813.5 KB│0.9540.888
No partner0│0.9260.830

In the revised PDF, the new Sec. 4.4 adds the latency (Table 7) and communication-cost (Table 6) studies, with the full sweeps in Appendices K and L.

Summary. We restated the contribution as a system plus an empirical finding, corrected the 16.0% sentence, filled the fusion cell, added latency and bandwidth studies, and revised joint fine-tuning. We would be glad to answer any follow-up questions.

Dear Reviewer kR6n,

Thank you again for your thoughtful feedback. We have posted our responses to your concerns and would appreciate your thoughts on whether they address the issues you raised. We are happy to provide any further clarification before the discussion period closes.

Reviewer_L4M6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies cooperative multi-agent multimodal end-to-end motion forecasting for autonomous driving. CooperTrace combines cross-agent BEV fusion, LiDAR-camera perception, query tracking, and explicit tracker history, and evaluates coverage, EPA, and conditional trajectory error. Its main insight is to address the mismatch between training and deployment history distributions and coverage loss from dropped tracks.

Critical strengths
  1. The problem is practically relevant. The paper incorporates tracker history and forecasting coverage into end-to-end motion forecasting rather than evaluating only successfully tracked targets.
  2. Tables 7–8 provide direct ablations of explicit history inputs and deployed-tracker history; Tables 4–5 report error, coverage, and EPA under sensor degradation.
  3. The experiments cover three cooperative datasets and evaluate detection, tracking, forecasting, and sensor degradation.
Critical weaknesses
  1. The main cross-method comparison is not directly comparable. Table 2 reports a 67% V2X-Seq minADE reduction, although UniV2X runs at 2 Hz and ours at 10 Hz. The methods also use different tracker-filtered target sets, and the CMP† predictor did not converge.
  2. The robustness evidence is limited. Tables 4–5 evaluate single-agent corruptions. They support coverage gains under the tested OPV2V simulations, but not broad cooperative or harsh-deployment robustness.
  3. The source of the history gain is not isolated. Tables 7–8 support explicit history and train-deployment history matching, but their single-agent, map-free, frozen-tracker design does not isolate the effect of the shared differentiable graph.
Questions for authors
  1. Can V2X-Seq ADE be recomputed at common time points and on a common target set, with the non-converged CMP† result clearly separated?
  2. What attribution of the history gain is supported by the current ablations?
2 comments

Thank you for the careful review and the concrete questions. We address each point in the order of the review. Changes in the revised PDF are in blue.

1. V2X-Seq comparability with UniV2X and the converged CMP (W1, Q1)

On the samples that both methods forecast (same target set, n = 4,612, same 2 Hz timestamps), the reduction is 65% instead of 67%. Retrained at 2 Hz, CooperTrace still reduces minADE@5s and minFDE@5s by over 40%, so the gap does not come from the frame rate.

│minFDE@5s ↓│minADE@5s ↓
Our 10 Hz model, n = 4,612││
UniV2X│5.314│2.953
CooperTrace│1.848│1.045
Reduction│65.2%│64.6%
Our model retrained at 2 Hz, n = 3,697││
UniV2X│2.862│1.716
CooperTrace│1.601│1.001
Reduction│44.1%│41.7%

CMP† did not converge because of a training bug (flipped box yaw), now fixed. Its minADE@5s on V2X-Seq drops from 4.281 to 1.875 m. Its V2X-Seq rows in the revised Table 2 are below (full table in the general response).

Method│minADE@5s ↓minFDE@5s ↓│Cov. ↑EPA ↑
CMP, submission (not converged)│4.2818.732│0.2220.094
CMP, converged│1.8754.308│0.2220.131
UniV2X│3.0475.496│0.318−0.033
CooperTrace│1.0191.873│0.8030.480

Both results are in the revised Table 2 and Appendix J.

2. Robustness beyond single-agent corruptions (W2)

We agree and added three experiments. (1) Cooperative models with every agent's sensors degraded (main-paper Tables 4 and 5, merged below). (2) The same corruptions on real-world data (appendix Table 20). (3) For harsh deployment, partner delays (Table 7 of the main paper) and compressed feature exchange (Table 6 below), both in Sec. 4.4. The finding of the submission still holds. When LiDAR degrades, the camera keeps coverage (0.631 vs. 0.327 under beam reduction, 0.579 vs. 0.187 with LiDAR missing, moderate severity), and camera corruptions leave fusion at its clean level.

│mAP_3D ↑ low│mAP_3D ↑ moderate│mAP_3D ↑ high│minADE@5s ↓│Cov. ↑│EPA ↑
Corruption│LL+C│LL+C│LL+C│LL+C│LL+C│LL+C
None (clean)│0.8760.863│0.8760.863│0.8760.863│1.3651.484│0.9110.888│0.4170.392
Fog attenuation│0.8410.827│0.7490.752│0.4600.573│1.3471.475│0.7810.809│0.3780.367
Beam reduction│0.8310.822│0.4600.576│0.0710.320│1.5341.776│0.3270.631│0.1260.238
LiDAR missing│0.6000.653│0.3280.507│0.1260.388│1.6331.820│0.1870.579│0.0650.211
Darkening│0.8760.868│0.8760.872│0.8760.876│1.3651.459│0.9110.872│0.4170.393
Camera missing│0.8760.868│0.8760.873│0.8760.877│1.3651.484│0.9110.887│0.4170.390

The two harshest real-world rows are below (moderate severity).

││mAP_3D ↑│minADE@5s ↓│Coverage ↑│EPA ↑
Dataset│Corruption│LL+C│LL+C│LL+C│LL+C
CooperScene│Beam reduction│0.2490.291│0.9560.938│0.2830.296│0.2090.203
CooperScene│LiDAR missing│0.2730.361│0.8130.953│0.2620.333│0.2040.228
V2X-Seq│Beam reduction│0.0760.111│1.5881.805│0.0540.140│0.0240.056
V2X-Seq│LiDAR missing│0.0560.155│1.8081.572│0.0560.218│0.0150.090

In Table 6, compressing the exchange from 18 MB to 13.5 KB costs only 1.5 points of mAP_BEV.

ExchangePayload│mAP_BEV ↑mAP_3D ↑
Dense fp1618 MB│0.9690.903
Top-k 0.5% + int850 KB│0.9550.899
Top-k 0.15% + int813.5 KB│0.9540.888
No partner0│0.9260.830

Appendix Tables 17–19 hold the single-agent, all-corruption and three-severity results.

3. Attribution of the history gain (W3, Q2)

Please see point 1 of the general response. The explicit history from the forecaster's own tracker is the main source of the gain, also in the full cooperative graph. For the effect of the shared differentiable graph, joint fine-tuning on live tracker outputs lowers minADE@5s over staged training from 1.036 to 1.019 m on V2X-Seq and from 1.405 to 1.374 m on OPV2V.

See Sec. 4.5 and Appendix I.

Summary. We recomputed the V2X-Seq comparison with a converged CMP, added cooperative, real-data and deployment robustness, and repeated the history ablations in the full cooperative graph. We would be glad to answer any follow-up questions.

Dear Reviewer L4M6,

Thank you again for your thoughtful feedback. We have posted our responses to your concerns and would appreciate your thoughts on whether they address the issues you raised. We are happy to provide any further clarification before the discussion period closes.

Comments on the paper

1 comment

We thank all reviewers (A: L4M6, B: kR6n, C: PPNL) for their careful reading and constructive feedback. Below we address the two concerns shared by the reviewers. The individual responses answer every point in full, and changes in the revised PDF are in blue.

1. Attribution of the gain: history, track query and BEV features (A, C)

Reviewer A noted that the history gain is not isolated in the single-agent, map-free, frozen-tracker design of the original Tables 7–8, and Reviewer C asked how a history-only forecaster performs and what the track queries and BEV features add. We repeated the ablations inside the full cooperative graph with map input.

The explicit history contributes most to the forecast. The history alone reaches a minADE@5s of 1.295 m (cooperative) and 1.794 m (single agent), against 5.908 m and 4.423 m for the track query alone.

The track query remains useful, but the way it is fused matters. Plain concatenation raises minADE@5s to 1.439 m (cooperative) and 2.020 m (single agent), so we added a gated fusion with a learned per-agent gate, trained with query dropout (p = 0.3). With the gate, EPA rises from 0.422 to 0.450 at identical coverage, as the query helps the tracks whose recent motion is a poor predictor of the next five seconds (Sec. 4.5).

Without BEV attention, the history-only minADE@5s rises from 1.794 to 1.902 m (single agent), so BEV features contribute about 0.11 m.

│minADE@5s ↓│Coverage ↑│EPA ↑
Cooperative, map (Table 9)│││
History only│1.295│0.911│0.422
Track query only│5.908│0.911│0.148
History + query (concatenation)│1.439│0.911│0.413
History + gated query│1.326│0.911│0.450
Single agent, map-free (Table 22)│││
History only│1.794│0.878│0.246
Track query only│4.423│0.878│0.172
History + query (concatenation)│2.020│0.878│0.210

2. Contribution and positioning, and the role of end-to-end training (B, C)

Reviewer B is right that explicit histories (VectorNet, MTR) and training on tracker output (PnPNet) exist in prior work, and we now cite them next to our contribution. What had not been done is to bring them into cooperative, multi-modal forecasting. Our contribution is both a system and an empirical finding.

  • A system. CooperTrace is the first framework to combine V2X cooperation, LiDAR–camera fusion, and perception-tracking-forecasting in one differentiable graph, where the forecaster is trained on the trajectories output by its own tracker (with its noise, missed detections and identity switches), as at test time. The "first" claim refers to this combination only.
  • An empirical finding. Query-centric cooperative systems lose accuracy because they discard explicit coordinates and are trained without realistic tracker noise.

In the submission, joint fine-tuning improved minADE@5s by only 0.027 m (original Table 6). We found that this experiment was not fully end-to-end, because the forecaster was still reading histories cached before joint training. Trained on the live tracker outputs, joint fine-tuning now improves over staged training in every metric on both real-world V2X-Seq and OPV2V, and Table 2 reports this model. Staged training also trains the forecaster by back-propagation on its own tracker's output, which is why it is already close.

│minADE@5s ↓│Coverage ↑│EPA ↑
Training (cooperative L+C)│V2X-SeqOPV2V│V2X-SeqOPV2V│V2X-SeqOPV2V
Staged│1.0361.405│0.7800.888│0.4710.426
Joint, cached histories (submission)│1.0871.472│0.7960.891│0.4660.418
Joint, live tracker outputs (revised)│1.0191.374│0.8030.893│0.4800.441

Revised Table 2 (CooperTrace rows are jointly fine-tuned, CMP on V2X-Seq is the converged run) is below.

│minADE ↓│minFDE ↓│
DatasetMethod│@1s@3s@5s│@1s@3s@5s│Cov. ↑EPA ↑
OPV2VCMP│0.3270.9411.833│0.5011.8403.982│0.7110.265
UniV2X│1.3012.4123.562│1.5853.6035.897│0.453−0.165
CooperTrace│0.1810.6251.374│0.2821.3073.261│0.8930.441
CooperSceneCMP│0.3300.5630.870│0.4070.8761.677│0.2190.190
UniV2X│1.0371.8822.979│1.1662.9335.243│0.151−0.091
CooperTrace│0.3230.5960.970│0.3620.9831.914│0.4620.313
V2X-SeqCMP│0.3540.9651.875│0.5201.9854.308│0.2220.131
UniV2X│1.0101.9313.047│1.1763.0415.496│0.318−0.033
CooperTrace│0.3660.6421.019│0.4091.0081.873│0.8030.480

We would be glad to answer any follow-up questions. Thank you!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.