TACO: Temporal Consensus Optimization for Continual Neural Mapping

Robotics: Science and Systems (RSS) 2026
1Nanjing University 2University of Illinois Urbana-Champaign 3University of California, Berkeley
Completed at the ICON Lab, led by Negar Mehr.
*Equal contribution

A replay-free framework for continual neural mapping that preserves reliable historical geometry with temporal consensus over neural implicit map snapshots.

Abstract

Continual neural mapping must update a scene representation from sequential RGB-D observations without catastrophically overwriting geometry learned earlier. Replay-based approaches reduce forgetting by storing historical observations, but this increases memory usage and can be undesirable in long-horizon robot deployments. TACO addresses this setting with temporal consensus optimization: historical neural map snapshots are treated as temporal neighbors, and the current map is optimized against them using parameter-wise importance weights. The resulting objective preserves parameters that strongly influence previously observed geometry while leaving uncertain or newly changed regions free to adapt.

Method Overview

Temporal Snapshots

TACO periodically stores compact snapshots of the neural implicit map instead of retaining old RGB-D frames.

Importance Weighting

Parameter importance is estimated from output sensitivity, identifying which parts of the map should be protected.

Consensus Optimization

The current map is optimized with a masked, importance-weighted temporal consensus term over past snapshots.

TACO temporal consensus pipeline
Graphical Explanation

Temporal consensus across evolving neural maps

TACO turns historical map states into temporal neighbors, preserving stable scene knowledge while allowing changed regions to update over time.

Results on Static Scenes

Static scene benchmark results
Static Scene Benchmark

Replay-free mapping remains competitive across reconstruction metrics

Across artifacts, holes, Chamfer distance, completion ratio, precision, and F1, TACO balances plasticity and stability without storing past RGB-D frames.

Results on Dynamic Scenes

Dynamic real-world scene results
Dynamic Real Scene

Adapting to scene changes without replay

TACO supports continual updates in dynamic real-world settings by protecting reliable map parameters and adapting uncertain regions as new observations arrive.

RSS Poster

BibTeX

@article{zhou2026taco,
  title={TACO: Temporal Consensus Optimization for Continual Neural Mapping},
  author={Zhou, Xunlan and Zhao, Hongrui and Mehr, Negar},
  journal={arXiv preprint arXiv:2602.04516},
  year={2026}
}