Skip to content
View sceneeeee's full-sized avatar

Block or report sceneeeee

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
sceneeeee/README.md

evenS

BEng student in Mechatronics and Robotic Systems at Xi'an Jiaotong-Liverpool University (XJTLU)
Expected graduation: 2028 · Based in Suzhou, China
Robotics Software · Embodied AI · Autonomous Navigation · Reproducible Evaluation

Room-Sign Navigation · NaVIDA × Habitat · ETPNav × MID360 · ROS 2 FishBot · All repositories

About

I am a third-year robotics undergraduate interested in connecting visual perception, language-grounded reasoning, and reliable robot software. My projects span RGB-based indoor navigation, vision-and-language navigation (VLN) model integration, geometric waypoint experiments, ROS 2 simulation, and reproducible engineering workflows.

I work primarily with Python, C++, PyTorch, OpenCV, ROS 2, Nav2, Gazebo, Habitat, Linux, Git, and Docker. These are tools used across different projects, not a claim of equal proficiency in every technology.

Current focus: RGB-first, human-in-the-loop room/sign-guided navigation. I am also seeking internships in robotics software, embodied AI engineering, perception, and simulation/evaluation.

Current Research

Room & Sign-Guided Indoor Navigation — RGB, OCR, reasoning and bounded guidance

A two-person research testbed investigating whether observed room numbers and directional signs can guide the search for an unknown indoor room without a pre-annotated target location. Its current handheld stage uses one RGB camera; depth, SLAM and Nav2 are not part of this perception/reasoning pipeline.

Verified progress

  • M0: Camera capture, RapidOCR, human review and data export are merged into main (PR #1).
  • M2.1–M2.4: Evidence-aware LandmarkMemory, room-number Rule baseline, constrained Qwen3-4B text planner, offline evaluation and JSONL event streaming are merged via PR #4 (merged on Nov 7, despite its historical WIP title/body).
  • M1 / M3: OCR-event ingestion, RGB door/sign proposals, short-lived tracks, bounded camera-relative suggestions and regression tests are integrated. PR #6 added validation/export of 21 human-reviewed instance records across 20 keyframes; the review includes incomplete frames.
  • Recorded-video evaluation: M3.1 runs detector → track → ROI OCR → M1/M2.4 → M3 on real prerecorded videos. M3.2 provides limited door-inspection suggestions, with motion_authorized=false. See recorded-video validation and M3.2 boundary.

My contributions: M2 reasoning and offline evaluation, rule/text-LLM comparisons, strict decision validation, JSONL replay/integration, and M3/M4 experiment planning. Collaborator-led contributions: M0 field capture/OCR and M1 perception, annotation and physical-door association. Integrated modules reflect joint project work, not sole authorship.

Boundary: Offline tests, candidate detection and prerecorded-video suggestions do not establish a live multi-door room-finding loop, trustworthy physical-door identity across sessions, collision-free robot motion, or verified goal discovery. M3.3 human-in-the-loop live acceptance is pending.

Completed Research and Engineering

NaVIDA × Habitat — real-checkpoint integration and separate systems showcase

SURF research period (Jul 2026): real model experiments. I developed the NaVIDA-side runtime and stateful HTTP policy service and contributed the evaluator-side HTTP adapter and grounded demo to the SURF team Habitat repository. Evidence includes:

  • Loading and running the official NaVIDA-3B checkpoint in 4-bit (NF4) on an RTX 4060 Ti 8GB; processing RGB observation history, language instructions and constrained action chunks.
  • Model/service → HTTP → Habitat execution integration, including a five-step engineering smoke test.
  • An 80-step, instruction/goal-grounded single-episode test with the real checkpoint. Execution completed, but NaVIDA did not reach the goal (success 0, SPL 0; about 8.620 m final goal distance). This is one historical episode, not an official VLN benchmark. The separate Oracle's success is not a NaVIDA result.

Post-SURF engineering closeout (Nov 15, 2026): deterministic Fake Policy. I consolidated a provenance-tracked evaluator subset and validated real Habitat rendering, localhost HTTP, action queues, recorded poses and video. The final 21-action / 21-request engineering showcase uses a deterministic Fake Policy, not the NaVIDA checkpoint, and demonstrates system execution rather than learned navigation. See validation evidence, provenance and labelled demo.

The SURF research line and this engineering deliverable are completed. No successful learned-navigation benchmark, model training, physical-robot deployment or fresh post-migration checkpoint replay is claimed.

ETPNav × MID360 — sparse-depth waypoint experiments (closed Nov 16, 2026)

Investigated adapting an ETPNav / Habitat waypoint proposal interface from dense RGB-D to MID360-inspired sparse geometry, while keeping downstream GraphMap and language/topological planning as stable as possible.

  • Reproduced the original ETPNav pathway, refactored waypoint providers and tested interface equivalence.
  • Implemented and compared Geometry V0, bounded NMS (BNMS), middle-range utility (MRU), and G-Min; audited how candidate geometry interacts with GraphMap.
  • Ran controlled sparse-depth zero-shot tests and activation-level failure analysis. The frozen waypoint predictor collapsed despite partly active depth-encoder features; simple feature-statistics alignment did not recover the predictor.
  • Reported comparisons on a fixed 20-episode development panel, not an official R2R-CE benchmark; see closeout and evidence pack.

Boundary: The sparse observations were deterministic approximations derived from simulated dense depth, not real Livox MID360 scans. Learned-adapter validation, Isaac Sim runtime and physical hardware deployment were not completed. This research line has been discontinued.

ROS 2 FishBot — mobile-robot simulation (completed)

Built a simulation-only ROS 2 Humble workspace spanning URDF/Xacro robot descriptions, Gazebo Classic, ros2_control differential-drive control, Nav2 map navigation, Python nav2_simple_commander waypoint examples, and a configurable patrol application with speech-service and camera-capture interfaces. Closeout. No physical FishBot deployment is claimed.

Bililive Recording & Upload Pipeline — reproducible software integration

Integrated existing recording/upload components with pinned upstream versions, SHA-256-verified patches, Docker Compose/systemd operations, SQLite-backed conservative cleanup, and Python/Node.js regression checks. The repository distinguishes historical production evidence from the current public prerelease, which has not passed a complete fresh-host production end-to-end acceptance. Validation boundaries.

Lecture-Grounded Tutorial Explainer — reusable ChatGPT Skill

Designed a source-first tutorial-explanation workflow that uses the learner's lecture materials, preserves notation, connects relevant figures to each question, and builds step-by-step solutions. It is a Skill specification, not a claim of a deployed standalone tutoring service.

Industry Experience

Package Testing Intern · Goertek Microelectronics Co., Ltd. · Aug–Sep 2025

Exposure to semiconductor packaging/testing operations and industrial quality-oriented workflows under team supervision.

Engineering Approach

I prefer explicit interfaces, controlled comparisons, tests and reproducible artifacts over unsupported performance claims. I distinguish unit tests, recorded-video experiments, simulation engineering demonstrations, single-episode results and standardized benchmarks. Research negative results and incomplete validation gates are documented rather than presented as successes.

Longer-term interests: deployable embodied navigation, 3D perception and robot software integration.

Pinned Loading

  1. room-sign-guided-navigation room-sign-guided-navigation Public

    Python

  2. ETPNav-MID360 ETPNav-MID360 Public

    ETPNav research on MID360-inspired sparse-depth adaptation, geometric waypoint proposals, GraphMap interface analysis, and predictor failure forensics in Habitat.

    Python

  3. navida-habitat navida-habitat Public

    NaVIDA–Habitat software integration: stateful HTTP inference, strict action parsing, Habitat evaluation, and deterministic simulation validation.

    Python

  4. ros2-fishbot ros2-fishbot Public

    ROS 2 Humble FishBot workspace with URDF/Xacro, Gazebo, ros2_control, Nav2, Python navigation demos, and autopatrol examples.

    Python 3

  5. bililive-record-upload-pipeline bililive-record-upload-pipeline Public

    Self-hosted Bilibili live recording and auto-upload pipeline with reproducible patches, conservative cleanup, and monitoring.

    Python

  6. VLN VLN Public

    Forked from Yangbadger222/VLN

    Python