Register-Token Sidecar for Data-Efficient Perception Pipelines

Independent Research Proposal
May 2026

A lightweight, non-invasive register-token sidecar concept for improving data triage, long-tail event mining, and upload efficiency in large-scale perception and autonomous-system pipelines.

Status: proposal and experiment scaffold. No production claims or benchmark results yet.

Independent work. Not affiliated with Tesla, Valeo, or the authors of DrivoR.

Summary

Many perception and robotics teams collect far more sensor data than they can fully upload, store, review, or label. At the same time, the most valuable data is often the hardest to surface: rare failures, ambiguous scenes, and long-tail edge cases.

This repository explores a low-risk sidecar design that runs in parallel with an existing perception or planning stack and produces compact learned scene summaries. The goal is to help teams decide which clips are most valuable to keep, upload, review, or label without modifying the main production model.

Why This Matters to Companies

This idea is aimed at organizations that operate data-hungry perception systems, including autonomous driving, ADAS, teleoperation, warehouse robotics, delivery robots, and other edge-AI fleets.

Potential value:

• improve useful-signal-per-GB under storage and bandwidth constraints
• surface rare or difficult examples that simple heuristics may miss
• reduce wasted upload, review, and labeling effort on low-value clips
• speed up offline search, triage, and dataset curation loops
• test a new data-efficiency layer without replacing the main production stack

Core Idea

Instead of treating every candidate clip equally, a sidecar can produce compact token-level summaries and lightweight scores for:

• novelty
• uncertainty
• long-tail value
• upload priority

These summaries can support:

1. ranking clips under fixed upload budgets
2. retaining compact metadata for fast retrieval
3. triggering full-resolution upload only when predicted training value is high
4. reducing review load for repetitive or low-information scenes

Proposed Architecture

Camera / Sensor Streams
        |
        v
Existing Perception or Policy Stack -----------------> Normal Logs / Telemetry
        |
        | parallel, non-invasive
        v
Register-Token Sidecar
        |
        +--> Compact scene summary tokens
        +--> Novelty / uncertainty score
        +--> Long-tail trigger score
        +--> Upload priority or triage signal

Hypotheses

H1: Long-Tail Mining

A compact register-token sidecar can improve recall of rare or difficult scenes at a fixed upload budget compared with baseline heuristic triggers.

H2: Payload Efficiency

Compact token summaries can help reduce data movement and storage cost by supporting smarter upload decisions while preserving useful training signal.

H3: Faster Curation

Token-level indexing can support faster offline search and triage than scanning raw clips alone.

Success Metrics

• rare-event recall at fixed upload budget
• useful examples found per GB uploaded
• false-positive rate of sidecar-triggered uploads
• retrieval speed for targeted hard-case search
• annotation triage time for selected clips

Offline Validation Plan

Week 1: Baseline Setup

• select one public dataset such as nuScenes mini or BDD100K
• define baseline triggers using scene rarity, lighting, motion, or proxy safety events
• label a small set of "interesting" clips for offline comparison

Week 2: Sidecar Prototype

• implement a lightweight token-summary or proxy representation
• compare clip ranking quality against heuristic baselines
• track false positives and false negatives at fixed upload budgets

Week 3: Efficiency Study

• compare raw-upload-first vs token-summary-first selection
• measure data efficiency, retrieval speed, and triage usefulness
• write up limitations, failure modes, and whether the idea survives kill criteria

Kill Criteria

Stop the approach if:

• rare-event recall does not beat simple heuristics
• false positives erase upload or storage gains
• token summaries do not preserve enough signal for useful triage
• sidecar complexity outweighs operational benefits
• results cannot be reproduced on at least two data slices or datasets

Repository Layout

This repository currently includes:

• README.md for the proposal and validation plan
• whitepaper.tex for the publication-style source document
• whitepaper.pdf for the compiled whitepaper
• experiments/ for future offline validation code and notes
• results/ for benchmark tables, plots, and analysis
• CITATION.cff and LICENSE for reuse and attribution

Planned additions:

• small baseline scripts or notebooks for offline evaluation

Related Work

• Driving on Registers (DrivoR)
• register-token methods for compact scene representation
• active learning and hard-example mining for perception systems
• fleet-scale data triage and shadow-mode data collection workflows

Collaboration

Feedback is welcome from researchers and engineers working on:

• autonomous systems
• robotics data engines
• computer vision infrastructure
• active learning
• long-tail scenario discovery
• dataset curation and triage

Citation

If you reference this work, please cite:

@misc{denialkhmbot_register_token_sidecar_2026,
  author = {Denialkhmbot},
  title  = {Register-Token Sidecar for Data-Efficient Perception Pipelines},
  year   = {2026},
  month  = {May},
  note   = {Independent research proposal},
  url    = {https://github.com/denial-web/register-token-sidecar-av}
}

License

MIT License for original text and code in this repository.

If future versions incorporate third-party code, the original license terms and attribution requirements should be preserved.