@inproceedings{Li2026e,

title = {GeoFlex: Geographic-Prior Flexible Adaptation of Per-Domain Diffusion Trajectory Planners to Held-Out Deployment Regions},

author = {Yun Li and Ehsan Javanmardi and Simon Thompson and Jiachen Gong and Qunli Zhang and Zifan Zeng and Shiming Liu and Peng Wang and Yidu Zhang and Manabu Tsukada},

year  = {2026},

date = {2026-11-09},

urldate = {2026-11-09},

booktitle = {Conference on Robot Learning (CoRL 2026)},

address = {Austin, Texas, USA},

abstract = {Cross-region deployment of a structured multi-domain diffusion planner reduces to a small-data fine-tuning problem rather than a per-scene routing problem. We test eight zero- or low-training routing schemes on three real-world held-out maps under a disjoint-lanelet protocol; none beats a trivial static rule that picks the geographically nearest training domain. A Fano bound and a split-conformal exchangeability rejection (p=0.0035) explain why no scene-feature router can break this static-prior ceiling, and a PAC-Bayes-certified von Mises-Fisher mixture, the strongest principled routing-class baseline, confirms the limit is architectural: any method confined to the linear span of the four per-expert weights is bounded by the static-prior fleet mean (15.85 vs. 13.74 m). Stepping outside this subspace breaks the ceiling. Joint replay, mixing target-region scenes with training scenes at 10:1 during fine-tuning, reaches 4.26 m fleet-mean on disjoint lanelets (-76% vs. frozen substrate) at +0.74 m in-region regression; the recipe replicates across all four nuPlan leave-one-city-out configurations including a -35% Singapore reduction under left-hand-traffic rule shift. As a methodological contribution we introduce variance-collapse forensics: a two-line diagnostic (within-scene K-sample variance plus cross-scene pairwise L2) that unmasks hypernetwork-conditioned LayerNorm and faithful TENT (test-time entropy minimisation) as near-stationary degenerate trajectories whose seemingly-strong fleet means would otherwise overstate the gain by an order of magnitude. The same backbone has passed real-vehicle closed-loop deployment under a production DPO LoRA (joint-replay weights are not yet on-vehicle), where adversarial-aware training lifts peak deceleration 3.4x without losing natural-data comfort.},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{nokey,

title = {HERO: Heterogeneous Evidential Robust Object-Level Collaborative Perception},

author = {Hao Si and Ehsan Javanmardi and Hanlin Wu and Manabu Tsukada },

year  = {2026},

date = {2026-09-08},

booktitle = {The 19th European Conference on Computer Vision (ECCV 2026)},

address = {Malmö, Sweden},

abstract = {Real-world collaborative perception agents typically deploy diverse sensor configurations and network architectures. While feature-level fusion methods address this heterogeneity by aligning feature spaces, they require collaborative training and large communication overhead. In contrast, object-level fusion sidesteps these constraints, as transmitted bounding boxes are directly usable across varying architectures. However, standard late fusion relies only on box coordinates and confidence scores, which provide insufficient information for robust association and leave the system vulnerable to localization noise and asynchronous delays. To address this, we propose HERO, an uncertainty-guided object-level collaborative perception method that equips proposals with evidential statistics and uses the uncertainty signals to perform robust association and fusion under pose noise and asynchrony. Experiments on OPV2V-H and DAIR-V2X show that HERO matches the SOTA feature-level fusion methods while reducing bandwidth by over 1000 times. Under localization error and latency scenarios, HERO further outperforms feature-level SOTA models.},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{Jiang2026b,

title = {CacheComm: A Motion-Aware Cache System for Cooperative Perception},

author = {Yuze Jiang and Ehsan Javanmardi and Bo Qian and Manabu Tsukada and Hiroshi Esaki},

year  = {2026},

date = {2026-09-06},

booktitle = {2026 IEEE 104th Vehicular Technology Conference (VTC2026-Fall)},

address = {Boston, MA, USA},

abstract = {Cooperative perception enables connected and automated vehicles (CAVs) to overcome the inherent limitations of single-agent sensing by sharing perceptual information through vehicular communications. However, existing intermediate-fusion approaches treat each transmission cycle independently, re-transmitting features for the entire scene regardless of how much has actually changed, resulting in significant bandwidth waste. We propose CacheComm, a motion-aware caching framework that exploits temporal redundancy in cooperative perception by utilizing an RSU as an intermediate feature cache. The RSU maintains an up-to-date feature map of its coverage area and uses lightweight motion detection to identify scene changes. It selectively updates the cached feature map and serve the cache to CAVs based on the motion map to save communication resources. Experiments show that our system saves 10% V2X communication load under sparse CAV and limited RSU range, and up to 77% with full RSU coverage. Our method can be easily expanded to cover more area to achieve more reduction on communication cost. Saved bandwidth can be used to transmit more critical information or further increase the perception range for a CAV.},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{Tao2026,

title = {TacitCollab: A Plug-and-Play V2X Safety Filter via Free-Form LLM Dialogue},

author = {Ye Tao and Manabu Tsukada and Hiroshi Esaki},

year  = {2026},

date = {2026-06-09},

urldate = {2026-06-09},

booktitle = {2026 IEEE 103rd Vehicular Technology Conference (VTC2026-Spring)},

address = {Nice, France},

abstract = {Autonomous vehicles increasingly rely on Vehicle-to-Everything (V2X) communication to coordinate in complex traffic scenarios. However, the existing approaches typically exchange predefined message formats such as waypoints or discrete actions, limiting their ability to handle exotic situations

that require nuanced negotiation. We propose TacitCollab, a V2X dialogue system that enables vehicles to coordinate collision avoidance through free-form natural language communication. TacitCollab leverages vision-language models to interpret traffic situations and engage in human-like dialogue, allowing vehicles

to share observations, negotiate priorities, and reach mutual agreements. The system is designed as a plug-and-play module that can be inserted into existing autonomous driving pipelines, accepting intentions in arbitrary formats and outputting updated intentions in the same format. We evaluate TacitCollab on 238 collision scenarios from the DeepAccident dataset. Results show a Collision Avoidance Ratio of 97.4%, with 84.2% of scenarios achieving coordinated outcomes where vehicles successfully

negotiate right-of-way. The average bandwidth of 1.17 KB per scenario demonstrates that text-based dialogue is practical for bandwidth-constrained V2X networks.},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{Li2026b,

title = {Aligning What Matters: Object-Centric and Structure-Aware Feature Alignment for Heterogeneous Collaborative Perception},

author = {Donghui Li and Ehsan Javanmardi and Manabu Tsukada},

year  = {2026},

date = {2026-06-09},

booktitle = {2026 IEEE 103rd Vehicular Technology Conference (VTC2026-Spring)},

address = {Nice, France},

abstract = {Collaborative perception enhances autonomous driving by enabling vehicles to jointly reason about the environment. Feature-level sharing is particularly attractive, as it preserves rich spatial information and allows uncertain regions to be jointly refined before object inference. However, in realistic deployments, vehicles often rely on heterogeneous perception backbones and preprocessing pipelines, resulting in substantial domain gaps that make feature fusion unreliable. Recent protocol-domain adapter frameworks mitigate this issue by mapping heterogeneous features into a shared latent space, but their alignment objectives are typically defined over the entire Bird’s Eye View (BEV) grid. Since foreground objects occupy only a small fraction of BEV space, such global reconstruction objectives are dominated by background regions and provide weak supervision for aligning detection-critical features. Moreover, purely global alignment does not explicitly preserve fine-grained spatial structures that are essential for accurate 3D object detection.

In this work, we revisit protocol-domain alignment from the perspective of detection relevance and structural consistency. We propose an object-centric and structure-aware adapter for heterogeneous collaborative perception. Specifically, we introduce BEV foreground masks derived from 3D bounding boxes to emphasize object regions during alignment, and incorporate a Structural Similarity Index Measure (SSIM)–based objective to enforce fine-grained structural consistency across heterogeneous feature domains. Importantly, the proposed method preserves the plug-and-play property of protocol-domain collaboration and requires no modification of the original perception backbones or fusion modules.

Extensive experiments under diverse heterogeneous configurations demonstrate consistent improvements over protocol-domain adapter baselines, achieving up to +0.19 AP@0.7 under challenging modality settings.



},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}

@inproceedings{Li2026c,

title = {V2XState: Intent-aware Spatial Basis Attention for Cooperative End-to-end Driving},

author = {Dongyang Li and Ehsan Javanmardi and Manabu Tsukada},

year  = {2026},

date = {2026-06-09},

urldate = {2026-06-09},

booktitle = {2026 IEEE 103rd Vehicular Technology Conference (VTC2026-Spring)},

address = {Nice, France},

abstract = {Cooperative autonomous driving relies on shared perception, yet most planning models implicitly treat spatial relevance as scene-driven and largely invariant to the ego vehicle's intent. We argue that spatial relevance for planning should be intent-conditioned: which regions matter depends on the specific maneuver being executed. This paper proposes V2XState, a framework that operationalizes this insight by using ego states and commands to modulate spatial attention over cooperative features. By incorporating kinematic-aware inductive biases through a state-gated basis attention mechanism, V2XState yields context-sensitive emphasis that aligns planning with current driving intent. We integrate this mechanism into a lightweight planning stack and observe that the resulting attention maps are both interpretable and intent-consistent. Experiments on cooperative driving benchmarks validate that intent-aware spatial attention leads to more maneuver-consistent planning, achieving a 8.2%/7.2% reduction in ADE/FDE relative to state-of-the-art baselines.},

keywords = {},

pubstate = {published},

tppubtype = {inproceedings}

}