Repository navigation
Releases
V0.3.4 · Landing image loading
V0.3.4 · Landing image loading
V0.3.4 is a patch release of V0.3.3.
Fix
Since V0.3.3 the Event Lab section follows the first screen of the landing page. Chrome starts native lazy images 1250 to 2500 px ahead of the viewport, so the section's images (about 1 MB: the lead image, four camera frames and seven sample thumbnails) loaded together with the page. On the outbound link of the public host (about 3 Mbps, measured on 2026-10-25 from Ningbo and from the United States over the public internet) they delayed the application chunks that "Open my sandbox" loads right after, and the start sheet took more than 5 s to appear. Landing images now get their address only when they come within 200 px of the viewport.
Verification
- D3 regression gate, light tier, on the release tree: contracts, web unit tests, lint, the public build, the K9 checks and the lab build passed. Version-pinned Python tests pass.
- Public end-to-end tests against https://anet4d.agentnetwork.org.cn, including
/lab/and the case "landing to an own sandbox in two clicks", passed 7 of 7. The/lab/publishing checks passed, except the connection check, which has no file of 12 MiB or more to test (known item F-131). - On a local build, the landing page requests no Event Lab image before the visitor scrolls, and all of them load once the section comes into view (desktop 1440 x 900 and phone 390 x 844).
V0.3.3 · ANet4D samples
V0.3.3 · ANet4D samples
V0.3.3 extends the public samples of the Event Lab, moves them forward on the landing page, completes the move of the repository and the sample dataset to the ANet4D name, and makes main the only branch of the repository.
Event Lab samples
The Event Lab (https://anet4d.agentnetwork.org.cn/lab/) now offers seven sample runs over five dev events in the world oryahovo-crop1. Every run was recorded in a Gazebo Sim lockstep run (Realistic physics tier, L2 observations, three UAVs) and offers the event replay, the RGB and thermal private views of each UAV in synchronized playback, and Rerun data inspection.
| Event | Event type | Policies |
|---|---|---|
crop-handoff-001 |
rendezvous_handoff | track_share, coverage_scan, static_guard |
crop-handoff-002 |
rendezvous_handoff | track_share |
crop-reid-001 |
reid_occlusion | track_share |
crop-search-001 |
target_search | coverage_scan |
crop-vehicle-001 |
vehicle_track | track_share |
The new runs are single-seed, single-repeat samples on public dev seeds. They are not ranked benchmark results, and no test event is published. The thermal videos are offline renders along the same flight, not observations delivered to the agents.
Public site
- The landing page shows the Event Lab section right after the first screen. The section leads with event replay and per-UAV private views and has a gallery of the seven sample runs.
- In the sample catalog of
/lab/, runs are grouped by event, and a click anywhere on a row opens the sample run. Before, only the run id was a link. - The World Hub links to the Event Lab.
Names and addresses
- Repository: https://github.com/ANetResearch/ANet4D. The former address redirects. README badges, CITATION, the documentation and the site use the new address.
- Sample media: served from the ModelScope dataset
inksong/ANet4D-samples(CC BY 4.0). The former datasetinksong/ANet-Drone4D-samplesis kept unchanged for earlier sample versions. - Product name: prose in the README, the documentation and the site uses ANet4D throughout. The legacy identifiers listed in ADR-223 keep their names, and so does the license text.
Branches
d2, d3-dev and m23/scene-target-foundations are merged into main and deleted; main is the only branch. Earlier releases are checked out by tag (v0.2.x for D2, v0.3.x for D3). Hosted CI, the D3 light gate and the pre-push hook now guard main (ADR-225 in docs/05).
Verification
- D3 regression gate, light tier, on the release tree: contracts, web unit tests, lint, the public build, the K9 checks and the lab build passed.
- On the CPU pool (dmax): the zero-impact replay comparison, the resimulation check and the import check passed. The D2 Python suite reported two tests that pinned the previous product version; both are updated in the release tree and pass. The performance step measured the targets budget case 17 to 27 % above the D2 baseline while other workloads shared the host. An interleaved A/B run of the same case on an idle machine measured V0.3.1, V0.3.2 and V0.3.3 within 1 % of each other (median of five runs each), so the release carries no measurable regression. The performance step is repeated in an idle window of the CPU pool (issue #136).
- Public end-to-end tests against https://anet4d.agentnetwork.org.cn, including
/lab/, passed 7 of 7 on the first V0.3.3 deployment. The/lab/publishing checks passed, except the connection check, which has no file of 12 MiB or more to test (known item F-131). - Every ModelScope link of the sample version (82 files) was checked for HTTP 200, a 206 response to a Range request, the CORS header, size and SHA-256.
V0.3.2 · ANet4D
V0.3.2 · ANet4D
V0.3.2 renames the product (formerly ANet Drone4D) to ANet4D: simulation infrastructure for collaborative embodied agents. It also merges the scene-target foundations branch and seals the current version. The released capabilities are the same as V0.3.1: the interactive sandbox, the D2 world runtime and the D3 Event Lab. Drones are the first class of entity that is implemented. Collaboration across entity types (ground vehicles, surface vessels, fixed sensors and software agents) is the forward direction of the project and is not part of this release (docs/06).
Name and addresses
- Product name: ANet4D. Releases up to V0.3.1 were published as ANet Drone4D. The rename is recorded in ADR-223 (
docs/05). - Public site: https://anet4d.agentnetwork.org.cn. The former domain drone4d.agentnetwork.org.cn returns a permanent redirect (301) to the same path on the new domain.
- Documentation: https://docs.agentnetwork.org.cn/anet4d/. Pages under the former
/drone4d/path redirect to the same page under/anet4d/. - Repository: the repository URL does not change in this release.
- Header images: the README header, the social image and the landing page header use the new name.
Agent SDK
- The package, import name and command line tool are now
anet4d(pip install anet4d,python -m anet4d.serve,anet4d-agent). The SDK version is 0.3.2. drone4dremains for one release as a deprecated alias.import drone4dresolves to the same modules and emits aDeprecationWarning, andpython -m drone4d.serveanddrone4d-agentprint a notice. Agents built for V0.3.1 run without changes.
Scene-target foundations (experimental, off by default)
The collaborator branch m23/scene-target-foundations is merged (ADR-224). It adds scene-target appearance, activity plans for pedestrians and life recipes. With the default configuration nothing of it is reachable:
- The server switch
AWR_TARGET_LIFE_ENABLE(default off) controls the three experimental routes, which return 404 when it is off. It also controls the sandbox templates T7 to T13 and the reset of the slow-task simulation clock. - The public build switch
VITE_AWR_TARGET_LIFE(default 0) removes the scenario designer and its texts from the public site. - Contracts move from 1.2.0 to 1.3.0 with new optional fields only (
docs/17§18.25). Layouts and the realtime subprotocol are unchanged. - Activity plans draw from a registered random stream (stream 24,
target_activity). Stream 23 is reserved for D4.
The branch also brings these sandbox fixes, which are active in all builds:
- The mobile layout no longer stalls at start-up when the viewport is hidden.
- Camera tracking releases correctly.
- Target configuration is cleared when the session epoch changes, and is retried while the REST service is starting.
- "Add target" is disabled until the configuration has loaded.
- The session bar can resume a sandbox.
Compatibility
These identifiers keep their names as legacy identifiers:
- protocol and schema identifiers (
drone4d.bench.wire.v1,drone4d.event.v1and others) - the
DRONE4D_*environment variables and the agent socket path - native Gazebo plugin names
- server install paths and service unit names
Recorded tapes, runs and submissions from V0.3.x remain valid.
Verification
- The D3 regression gate ran on the release tree (tree
95a3f4890b83). Contracts, web unit tests, lint, the public build, the K9 checks and the lab build passed. On the CPU pool, the full D2 Python suite, the zero-impact replay comparison (relay_batch --determinism), the resimulation check and the import check passed. - The performance step was not judged: the shared CPU pool host ran at about 0.62 load per CPU during the gate, above the 0.5 limit. The changes in this release do not touch the simulation hot path; they rename the product, and the new code is behind switches that are off by default. The performance measurement is carried as an open item and will be repeated when the host is idle.
- Demo build parity is rebaselined to
v0.3.2. The public build differs from V0.3.1 by the new name and the sandbox fixes listed above.
Assets and attribution
The world model "oryahovo - crop #1 - metashape texture problems" is by Mitko (https://sketchfab.com/bg3d) and is used under CC BY 4.0, modified. Sample media, slices and the Rerun Web Viewer module are served from the ModelScope dataset inksong/ANet-Drone4D-samples (CC BY 4.0). The dataset keeps its name in this release.
Known issues
- The public host has limited outbound bandwidth (about 3 Mbps, measured on 2026-10-25 from Ningbo and from the United States over the public internet), so large media is served from ModelScope.
- The redirect pages under the former documentation path cover HTML pages only.
- Building from the
v0.3.2tag on a clean checkout:make setupfails becauseapps/webstill pins@awr/contracts1.2.0 (the workspace package is 1.3.0), andmake buildreports a false finding in the production bundle scan. Both are fixed onmain(9898185,6f55081); build frommainor apply these two commits. The deployed public build is not affected.
V0.3.1 · Event Lab on the public site
V0.3.1 · Event Lab on the public site
V0.3.1 makes the D3 Event Lab part of the main public site at https://drone4d.agentnetwork.org.cn. The main build of the site is now built from this release; earlier V0.3.0 kept the D2 build as the main site and served the Event Lab only at /lab/.
Public site
- Event Lab entry: the landing page links to the Event Lab from the header, the first screen and a new section, "Event Lab · UAV camera views".
- UAV camera views: the new section shows frames that the UAVs of a sample run actually received. They are rendered server-side from the GLB mesh world in the Gazebo Sim lockstep run, using the
rgb1080-rsrolling-shutter RGB camera and thetherm512thermal camera (offline render, white-hot palette). Each UAV sees only its own camera. - Sample run page:
/lab/samples/<run>now opens at an airborne moment instead of t = 0, when the fleet is still on the takeoff pad. A?t=query parameter selects any other time. - Loading bar: the loading bar is now a single determinate bar showing stage, percentage, megabytes loaded against the known total, smoothed network throughput and the estimated time remaining. The same bar is used when switching worlds and when resuming from power save. The previous five-segment bar could appear as two separate bars when shader warm-up finished before the first-screen point cloud.
Versions
- Product version 0.3.1 (
pyproject.toml,package.json,CITATION.cff, BibTeX). - The build-parity check of the regression gate (
make demo-parity) now compares againstv0.3.1, because the public main build is the build of this tag.
Assets and attribution
The world model "oryahovo - crop #1 - metashape texture problems" is by Mitko (https://sketchfab.com/bg3d) and is used under CC BY 4.0, modified. Sample media, slices and the Rerun Web Viewer module are served from the ModelScope dataset inksong/ANet-Drone4D-samples (CC BY 4.0).
Known issues
- Network throughput is under-reported on very slow links. The browser reports a download only once it has finished, so transfers still in progress are not counted.
- The public host has limited outbound bandwidth (about 3 Mbps measured on 2026-10-25 from Ningbo and from the United States over the public internet), so media is served from ModelScope.
V0.3.0 · Event Lab
V0.3 adds the Event Lab: mesh worlds, scripted events, Gazebo Sim lockstep physics, multi-UAV camera, thermal and LiDAR sensing, and an internal event-level observation benchmark. Sensor generation for the Event Lab requires a Linux x86_64 server with an NVIDIA GPU (EGL, driver 535 or later, at least 8 GB of video memory). The Interactive Sandbox (V0.2) is unchanged, and its server still needs no GPU.
Added
- Mesh world import: GLB scenes become metric World Packages with calibration and levelling, collision and ray-cast geometry, DSM and DTM, a Recast/Detour navigation mesh, anchors and browser LODs. Builds are deterministic and cross-checked on two machines.
- Event runtime: the
drone4d.event.v1event language with three compile gates and a byte-reproducible event tape. The first event set covers re-identification under occlusion, rendezvous and handoff, vehicle tracking (OpenStreetMap lane graph, car-following and junction yielding) and search for a target at an unknown location. - Physics tiers: Gazebo Sim lockstep (the default tier and the only source of formal results) and the FleetSim fast tier, with a tier-consistency comparison.
- Sensor generation: RGB cameras including rolling shutter and 1080p multi-UAV output, a thermal camera, mesh LiDAR, IMU and GNSS, generated at the same instant as the poses and the event tape.
- Agent observation SDK: a lockstep environment with observation levels L0 to L3, decision latency, isolated agent containers, a Python SDK and baseline policies.
- Benchmark runtime: lossless recording with retention, an evaluator reporting event capture, event understanding, active observation efficiency and collaborative efficiency with confidence intervals, internal submissions, and an internal multi-lane leaderboard. The leaderboard is internal and not offered on the public site.
- FAST-LIVO2 reconstruction comparison against mesh ground truth.
- Rerun-based data inspection with a shared timeline, and dataset export.
- Public read-only Event Lab samples at https://drone4d.agentnetwork.org.cn/lab/: event replay, per-UAV RGB and thermal videos synchronised to simulation time, and data slices. Sample data is published on ModelScope (dataset
inksong/ANet-Drone4D-samples, CC BY 4.0).
Changed
- The public site keeps the V0.2 Interactive Sandbox as its main build; the Event Lab samples are a separate static build under
/lab/. - Large sample files and the Rerun Viewer module are served from ModelScope, so downloads do not compete with the live sandbox for the demo host's bandwidth.
- README, landing page and citation metadata describe V0.3 as released.
Fixed
- D2 behaviour is unchanged: the regression gate confirms identical event logs, statistics and replay frame digests against the V0.2 baseline, and a byte-identical V0.2 demo build.
Known limitations
- The thermal videos in the public samples are rendered offline along the recorded flight; the V0.3 observation tier used for the samples does not deliver thermal frames to agents.
- Full-resolution sample videos are not yet published.
- Some acceptance evidence is scheduled for re-measurement in the first V0.3.x update; see the delivery summary in
docs/impl/D3-交付总结.md. - Planned for V0.3.x: the remaining Event Lab scope (event editor, larger worlds and event families, further sensor and appearance work).
V0.2.2 · Copy review of the release surface
Patch release of V0.2 (Interactive Sandbox). It contains the copy review of the release surface and the fixes released in v0.2.1. Server requirements are unchanged: Linux x86_64 with 4 CPU cores and no GPU.
Added
- The README states the runtime prerequisites of each capability and a requirements table: the Interactive Sandbox (V0.2) needs no server GPU; sensor generation in the Event Lab (V0.3, in development) requires an NVIDIA GPU with driver 535 or later.
Changed
- README and README.zh-CN follow one structure: capabilities with delivery and runtime prerequisites, requirements, quick start, documentation, license and attribution, and citation. Performance and scale figures state their hardware and measurement method.
- The product name is written as ANet Drone4D in the README, CITATION.cff, the about dialog and the landing page; the repository identifier remains in URLs only.
- Interface strings name the 3D view "三维场景视图" (Scene View): the "在三维场景视图中打开" button on the jobs page and the narrow-window hint.
- The English title of the public demo uses sentence case ("ANet Drone4D public demo").
- The landing page states that the Event Lab's sensor generation requires a server GPU.
- CITATION.cff and the BibTeX entries carry version 0.2.2.
Fixed
make lintno longer reports public model identifiers from the local configuration as leaked credentials (REL-03).
Released in v0.2.1:
- The Python SDK's queue long polls use their own read timeout (#14).
- Session usage reports nests and active tasks (#7).
- The application chunk is retried on a failed load, with a reload offered when it still fails (#9).
- The add-vehicle dialog disables nests that are full or incompatible (#8).
- HOLD and signal-delay states are judged by each vehicle's own sample interval, which removes false STALE reports (#6).
Known limitations
- The known issues listed in the v0.2.0 release notes still apply; see docs/impl/D2-验收报告-第2轮.md.
- The banner image (
docs/media/drone4d-banner.jpg) still shows the repository identifierANet-Drone4Das artwork.
V0.2 · Interactive sandbox and relay surveillance
ANet Drone4D V0.2 extends the V0.1 multi-agent 4D simulation platform with an interactive sandbox, quantified target perception, nest-based orchestration and continuous relay surveillance. The public demonstration at https://drone4d.agentnetwork.org.cn runs this release on CPU-only infrastructure.
Highlights
- Interactive sandbox. Each visitor session runs an isolated simulation core with its own fleet, targets and tasks. Admission control, a visible queue and an energy-saving mode reclaim server resources after inactivity.
- Target management and quantified perception. Six target classes (person, machine, dog, vehicle, object and unmanned aircraft) carry size, thermal emissivity, periodic acoustic signatures and sensitivity volumes. Detection follows Johnson/TTPF criteria for daylight, low-light and thermal sensors, with acoustic, visual, thermal and radar field views.
- Mission orchestration. Seven task types (point watch, polyline patrol, multi-aircraft area patrol, GSD-constrained area scan, perimeter patrol, area guard and relay watch) are planned by a greedy allocator with nest battery scheduling.
- Relay surveillance. Continuous 7x24 monitoring with zero-gap handover keeps observers outside each target's sensitivity range.
- Aircraft catalog. P600 with MID-360, small and heavy multirotor, long-endurance fixed-wing and VTOL presets, with daylight, near-infrared, thermal, radar and acoustic payloads.
- Open interface. REST API at
/api/sandbox/v1with an OpenAPI description and a Python SDK (tools/sdk/python) for fleet control and sensor and status access. - Seven world basemaps. A synthetic city and six UrbanScene3D cities, used with the authorization of the rights holders.
- Device coverage. Boot screen with version and progress, tablet command layout and simplified phone layout.
Acceptance
Second acceptance round: 44 of 55 criteria pass, 9 do not, and 2 require the public deployment and will be measured on the live site. The 24-hour soak (D2-AC-39) ran its full duration with no restarts, no out-of-memory events and memory growth of 0.48 MB/h. Full report: docs/impl/D2-验收报告-第2轮.md.
Known issues
- Sandbox capacity on the public host. With four full sessions, the sandbox CPU group saturates and per-session real-time factor drops to about 0.58 (D2-AC-02, 04, 07, 08, 55). The public deployment caps concurrent sandbox sessions at two; further visitors queue.
- Relay gaps in daytime. Random-walk targets briefly lose line of sight to the on-station observer, producing short monitoring gaps (D2-AC-19, 20, 35).
- Software rendering. Without a hardware GPU, the synthetic-city default view renders at a median of 50 ms per frame (D2-AC-30). Hardware-accelerated browsers are unaffected.
ANet Drone4D V0.2 在 V0.1 多智能体四维仿真平台的基础上,增加交互式沙盒、目标感知量化、机巢编排与持续接力监视。公开演示站 https://drone4d.agentnetwork.org.cn 已运行本版本,部署于无 GPU 的服务器。
主要能力
- 交互式沙盒:每位访客的会话拥有独立的仿真核心、机队、目标与任务;准入控制、可见排队与空闲节能模式按需回收服务端资源。
- 目标管理与感知量化:六类目标(人员、机械、犬只、车辆、物体与无人机)具备尺寸、热发射率、周期声学特征与敏感空间;可见光、微光与热成像按 Johnson/TTPF 判据判定,并提供声学、视觉、热与雷达场视图。
- 任务编排:七类任务(定点监视、多点巡线、多机区域巡查、指定 GSD 全域扫描、巡边、区域警戒与接力监视)由贪心分配器与机巢电池调度统一规划。
- 接力监视:7×24 小时持续监视,零间隙交接,观测站位保持在目标敏感范围之外。
- 机型目录:P600 与 MID-360、小型与重载多旋翼、长航时固定翼与垂起固定翼,可配置可见光、近红外、热成像、雷达与声学载荷。
- 开放接口:
/api/sandbox/v1REST 接口、OpenAPI 文档与 Python SDK(tools/sdk/python),可控制机队并读取传感器与状态。 - 七个世界底图:合成城市与 UrbanScene3D 六座城市,后者经数据集权利方授权使用。
- 多端适配:启动屏显示版本与载入进度;平板提供指挥布局,手机提供简化布局。
验收
第二轮验收:55 项中 44 项通过、9 项未通过,2 项需在公开站部署后实测。24 小时浸泡测试(D2-AC-39)按原标准完整运行,无进程重启、无内存溢出,内存增长 0.48 MB/h。完整报告见 docs/impl/D2-验收报告-第2轮.md。
已知问题
- 公开站沙盒容量:4 个满载会话时沙盒 CPU 组饱和,单会话实时因子降至约 0.58(D2-AC-02、04、07、08、55)。公开站将并发沙盒会话上限设为 2,其余访客排队。
- 白天接力间隙:随机游走目标短暂脱离在岗机视线,产生短时监视间隙(D2-AC-19、20、35)。
- 软件渲染:无硬件 GPU 时,合成城市默认视角的帧时间中位数为 50 ms(D2-AC-30);硬件加速浏览器不受影响。