blind-assist
BlindAssist is an Android showcase research prototype for goal-driven visual assistance. It combines camera perception, risk logic, and concise guidance to demonstrate measurable effects in clearly stated controlled conditions. It is not a certified mobility or safety product.
App v10.15.1 opens the original home without starting perception. Start assist explicitly starts the A+LOCAL experimental mode with the frozen research classifier running offline on the phone. Its nominal camera/ToF registration is not physical calibration; simulation results do not establish hardware accuracy. The original ToF mode remains available. See hardware guide.
The obstacle-perception research goal is 盲杖互补的类别无关前视障碍感知 (cane-complementary, class-agnostic forward obstacle awareness). Priorities are forward walls and large obstructions, body/head protrusions, suspended hazards, and multi-height poles, with useful direction and coarse range evidence under limited compute. The wearer chooses movement; this is not autonomous avoidance. Very low obstacles remain secondary compatibility evidence, not a headline optimization target. Knee-height hazards are not discarded. Earlier warning and dynamic-obstacle coverage are objectives requiring temporal validation.
The current research design combines base near-obstacle evidence with optional ground-relative enhancement. Missing ground may remove height refinement but must not erase an already observed near obstacle. This design and the new task priorities are research direction, not a claim that all capabilities are already deployed or validated. See the current perception route.
Start here
- 默认入口(v10.15.1):打开 App 停留原首页;点击“开始辅助”才运行相机与 10 Hz 8×8 ToF 的 A+LOCAL,点击“结束辅助”回首页并停止会话。支持基础 ToF 切换、事件提醒、最近 20 秒留证、静音回放与分路降级。
- 中文一页状态(系统、指标、当前问题): docs/PROJECT_STATE.md
- Current research decision: docs/CURRENT_DECISION.md
- L10-R0 paused / retained evidence: research/active/l10-r0/CURRENT.md
- Forward obstacle perception / retained DTR history: research/active/dtr-r0/CURRENT.md
- Code ownership: docs/CODE_MAP.md
- Documentation map: docs/README.md
- Historical lookup: docs/history-index.md
Forward obstacle perception is the sole advancing research line. L10 is paused;
its results and reproduction paths are retained. Closed
experiments are summarized in experiments/index.jsonl and remain recoverable
from the remote tag archive/pre-agent-surface-2026-08-26 or the terminal
commits listed in docs/history-index.md.
Workstation profiles
Copy the local template once and edit only machine-owned values:
Copy-Item config/local.example.toml config/local.toml
pwsh -NoProfile -File tools/ba.ps1 setup base
pwsh -NoProfile -File tools/ba.ps1 doctor base
Profiles are independent: base, research-l10-r0, research-dtr-r0,
android, device, and export. Research setup does not install or probe
Android tooling.
pwsh -NoProfile -File tools/ba.ps1 setup research-dtr-r0
pwsh -NoProfile -File tools/ba.ps1 doctor research-dtr-r0
pwsh -NoProfile -File tools/ba.ps1 smoke research-dtr-r0
The Codex desktop environment exposes the same setup and common actions through
.codex/environments/environment.toml.
Android
Run Gradle only through the repository wrapper:
pwsh -NoProfile -File scripts/run_android_gradle.ps1 :app:assembleDebug
Module boundaries are stable: :app owns the shell and assets,
:feature:assist runtime coordination, :core:assist pure risk logic,
:core:vision detection, :core:device Android adapters, and :core:ui state
and rendering.
Evidence boundary
Synthetic, replay, curated Development, device, and natural evidence are named
separately. UNKNOWN and NOT_EVALUABLE are not negative evidence. A focused
demo result is never presented as universal real-world or safety performance.
Protected final evaluations use formal research governance.
License
See CONTRIBUTING.md, GOVERNANCE.md, the model card, maintainer automation, the threat model, LICENSE, THIRD_PARTY_NOTICES.md, and SECURITY.md.
