AUTONOMOUS DRIVING · GEOMETRIC FOUNDATION MODELS

Drive3R

Block-wise Streaming Geometric Foundation Models
for Autonomous Driving

Geometry that goes the distance.

Tianyi Yang1,2Haojun Yang1Jipeng Xiao1Xuesong Chen2Shaoshuai Shi2Li Jiang3Zhipeng Zhang1,*

1 AutoLab, School of Artificial Intelligence, Shanghai Jiao Tong University

2 Voyager Research, Didi Chuxing · 3 The Chinese University of Hong Kong, Shenzhen

* Corresponding author

Paper Code
01 / THE IDEA

Local context.
Long-range consistency.

Driving logs do not end after a short clip. Their geometry should hold together for the whole journey.

Fully bidirectional models struggle with the memory cost of long sequences. Per-frame streaming loses local future context and accumulates drift. Fixed chunks preserve local reasoning, but fragment the global scene.

Drive3R brings the strengths together: bidirectional attention within each block, causal feature propagation across blocks, and confidence-aware re-anchoring when the streaming state becomes unreliable.

Read the abstract

High-fidelity closed-loop simulation for autonomous driving increasingly relies on dense multi-view 3D reconstruction. While recent feed-forward geometric foundation models have revolutionized per-scene optimization, applying them to continuous, long-horizon driving sequences is computationally intractable due to GPU memory limits and the quadratic complexity of global attention. Existing scaling strategies are fundamentally limited, as causal streaming suffers from compounding scale drift and trajectory deviation, while rigid “chunk-then-align” heuristics fragment global context and produce trajectory discontinuities. To address these limitations, we introduce Drive3R, a block-wise streaming geometric foundation model. Its block-wise causal attention resolves the scalability bottleneck of fully bidirectional attention while avoiding the loss of local future context in purely causal streaming. Its confidence-aware re-anchoring further mitigates long-range drift by detecting unreliable streaming states, re-initializing the KV cache, and stitching refreshed segments through SE(3) alignment. Extensive experiments demonstrate that Drive3R establishes a new state-of-the-art on nuScenes, KITTI, and Waymo. Specifically, on nuScenes, it improves over the recent DVGT-2 baseline from 0.1806 to 0.1081 in AbsRel, from 0.6750 to 0.5344 in point-cloud Accuracy, and from 1.5812 to 0.9092 in ATE.

nuScenes · Depth AbsRel ↓0.108140.1% lower than DVGT-2
nuScenes · Point-cloud Accuracy ↓0.534420.8% lower than DVGT-2
nuScenes · Trajectory ATE ↓0.909242.5% lower than DVGT-2
02 / HOW IT WORKS

Think in blocks.
Keep the bigger picture.

One model for metric geometry,
ego-motion, and streaming reliability.

Drive3R pipeline: intrinsic-aware image encoding, a spatio-temporal decoder with streamed KV memory, and prediction heads for geometry, ego pose, metric scale, and confidence.
THE ARCHITECTUREMulti-camera image blocks → intrinsic-aware encoder → spatio-temporal decoder → dense geometry and ego-motion.
01

Block-wise causal attention

Frames and camera views inside a block reason together bidirectionally. Across blocks, a causal KV cache carries historical features forward without recomputing the entire sequence.

Local bidirectional contextCross-block memory
02

Confidence-aware re-anchoring

A block-level confidence head monitors the streaming state. When confidence drops, Drive3R resets the cache and uses overlapping poses to align a refreshed segment into the global frame with SE(3).

Adaptive cache resetSE(3) alignment
Comparison of offline, per-frame streaming, chunk-then-align, and Drive3R block-wise streaming attention patterns.
FOUR PARADIGMSDrive3R preserves local future context while maintaining a continuous streaming memory.
03 / THE EVIDENCE

Better geometry.
Measured across benchmarks.

Full experimental details

Dense geometry & ego-motion

15 frames · full camera views

nuScenes results under the 15-frame, full-view setting
MethodAbsRel ↓δ < 1.25 ↑Accuracy ↓Completeness ↓AUC@30 ↑ATE ↓
DVGT-20.18060.72850.67500.623085.081.5812
Drive3R Ours0.10810.90770.53440.537486.920.9092

Drive3R advances depth estimation, dense scene geometry, and ego-motion on nuScenes, reducing depth error by 40.1% and trajectory error by 42.5% versus DVGT-2.

Bold: best result. * Baselines aligned to metric scale.

41.7average ATE ↓

Complete driving logs, without losing the thread.

Drive3R reduces average trajectory error by 44.8% compared with DVGT-2 across 11 complete KITTI driving sequences.

KITTI / 11 SEQUENCES
04 / A CLOSER LOOK

Structure that stays sharp.

Road surfaces, object boundaries,
and long-range trajectories.

Comparison of local scene geometry and object boundaries reconstructed by streaming baselines and Drive3R.
DENSE RECONSTRUCTIONDrive3R reconstructs cleaner local geometry with sharper object boundaries.
Estimated trajectories compared with ground truth on KITTI sequences 09 and 10.
LONG-SEQUENCE EGO-MOTIONKITTI sequences 09 and 10. Estimated trajectories are shown in teal; ground truth is shown as dashed gray curves.
BUILD ON DRIVE3R

A foundation for
the road ahead.

For dense metric reconstruction and
long-horizon autonomous driving research.

Read the paper

BibTeX

@misc{yang_drive3r,
  title = {Drive3R: Block-wise Streaming Geometric
    Foundation Models for Autonomous Driving},
  author = {Yang, Tianyi and Yang, Haojun and
    Xiao, Jipeng and Chen, Xuesong and Shi, Shaoshuai
    and Jiang, Li and Zhang, Zhipeng},
  note = {Research manuscript},
  url = {https://arxiv.org/abs/2610.15112}
}