Overview

Positioning

ANet4D is simulation infrastructure for collaborative embodied agents. It is built to study how teams of agents decide together where, when and how to sense and act in a dynamic real-world place: multi-UAV fleets today, and in the forward direction also unmanned ground vehicles (UGV), unmanned surface vessels (USV), ground robots, fixed sensors and software agents in one experiment. Experiments are meant to be reproducible and auditable, so that one team can rerun, inspect and replace each part of another team's setup.

Robot simulators usually start from the vehicle and give it a hand-built map; digital twins usually show a place but cannot run missions in it. ANet4D starts from the place: a captured or reconstructed point cloud becomes a versioned World, the environment field, the drone fleet and their agents all run on that World under one simulation clock, and a browser watches and operates it from anywhere. The whole chain is:

Reality → Reconstruction → World → Environment → Simulation → Agent

"4D" means three spatial dimensions plus time: drones, weather and the timeline share one simulation clock, so any moment can be paused, stepped, recorded and replayed.

Four design principles:

  • The World is the core. Capture, reconstruction and simulation converge on one World Package: geometry for physics and queries, a point cloud for display, semantics and no-fly zones, anchored to a real place.

  • The browser views the world, the server computes it. Physics, safety, planning, perception and environment run on a Linux server without a GPU; the browser only streams and draws.

  • The environment is a 4D field. Wind, turbulence, visibility and precipitation form one field E(x, y, z, t); the same numbers act on the airframes and paint the sky.

  • An entity is a Physical Agent. Each drone (the entity type delivered today) publishes what it can do; other agents discover it, request a quote and delegate tasks. Every flight command still passes the simulator's admission and safety checks.

The ANet4D sandbox in the synthetic demo city ANet Synthetic City: world, layer and weather panels, fleet list, timeline and performance HUD

A real screenshot of V0.1 (D1) in the synthetic demo city ANet Synthetic City (an 8-core machine without a GPU, software rendering).

Current version

Item

Content

Version

0.3.2 (V0.3.2 · ANet4D): the V0.3 Event Lab sealed with the tag v0.3.0 on 2026-10-25, the rename to ANet4D and the experimental scene-target foundations; each GitHub Release lists its known issues

Public instance

anet4d.agentnetwork.org.cn: the shared show, one sandbox per visitor and read-only Event Lab samples, no installation or sign-in

Server

Interactive Sandbox: Linux x86-64, Python 3.12, no GPU required; the public instance runs on a 4-core server without a GPU. Event Lab: Linux x86-64 with an NVIDIA GPU for sensor generation

Browser

Any desktop, tablet or phone browser with WebGL2; image quality and frame rate depend on the visitor's GPU

License

ANet Open Source License (a modified Apache-2.0), see Citation and license

Capabilities and delivery status

Status labels:

  • Delivered: released in a version and accepted, or listed as a known issue.

  • In development: design finalized and scheduled in the current phase, with interfaces and acceptance criteria defined; not yet released.

  • Research: under evaluation or prototyping; whether it enters the product depends on the evaluation.

Capability

Description

Status

World Package v1

Point cloud normalization (units, up axis, leveling, north, origin, normals, ground model), Potree 2.0 compatible octree with ANET_Q16 encoding, DSM and DTM, semantics and zones; validate --deep

Delivered V0.1

Seven worlds

The synthetic demo city ANet Synthetic City and six UrbanScene3D cities (authorized by the dataset rights holders for this project's deployments)

Delivered V0.1 / V0.2

City-scale point cloud streaming

One Range request per octree root for the first screen; the APH selector and CAS controller regulate the point budget across 7 quality tiers in a closed loop

Delivered V0.1

4D environment field

Wind (height profile and turbulence), visibility and precipitation, 12 weather presets; simulation calendar, sun position, illuminance and wavelength-dependent extinction

Delivered V0.1 / V0.2

Fleet on one clock (FleetSim)

1 to 1,000 drones on a shared 250 Hz clock; PX4-style cascade compiled with numba; 1,000 drones stay real time on about half a CPU core

Delivered V0.1

Safety and missions

14-state flight FSM, geofence, energy-aware return to home, separation guard; 8 mission generators

Delivered V0.1

Timeline, recording and replay

Pause, step, rate ×0.25 to ×10; MCAP recording with byte-identical replay (local runs; not offered on the public instance)

Delivered V0.1

Agent collaboration

In-process Mock ANet: capability discovery, quotes, delegation and receipt verification (search and rescue scenario S3)

Delivered V0.1

Interactive sandbox

One isolated session and sim-core per visitor; admission, queueing, renewal, snapshots and world switching

Delivered V0.2

Targets and perception quantification

Six target classes with sensitivity ranges; detection, recognition and identification unified by the Johnson criteria and TTPF; per-pair perception readings and field views

Delivered V0.2

Nests, tasking and relay

Six task types and area guard, greedy allocator behind the Allocator interface, GSD scan planning, 7x24 relay monitoring and sustainability assessment

Delivered V0.2 (short daytime relay gaps are a known issue)

Open interface and Python SDK

Sandbox REST v1, the awr.rt.v1 realtime protocol, OpenAPI 3.1 and a single-file SDK

Delivered V0.2

Airframe catalog

The P600 digital twin (MID-360 configuration) and four generic reference airframes, including fixed-wing and VTOL

Delivered V0.2

Startup, capacity and devices

Startup screen with version and progress, queues for viewer seats and sandbox slots, power saving, phone and tablet layouts

Delivered V0.2

Mesh worlds

Photogrammetric GLB meshes with metric scale, true north and leveling calibration, and derived geometry, navigation and quality layers

Delivered V0.3

Event simulation and active observation benchmark

Compare the observations of different fleet policies under the same world and event; event DSL and event tape, evaluation and an internal leaderboard

Delivered V0.3

High-fidelity physics tier

The gz-sim lockstep tier as the physics tier for official scores; FleetSim stays as the fast tier

Delivered V0.3

Sensor generation service

Mesh rasterization cameras (RGB, depth, instance, semantics, rolling shutter, thermal), mesh LiDAR with the Livox MID-360 scan pattern, and IMU synthesis

Delivered V0.3

Agent SDK and baselines

The anet4d Python SDK, the agent wire protocol over Unix sockets, isolated agent containers and three baseline policies

Delivered V0.3

Scene-target foundations

Target appearance, activity and population bindings as additive optional contract fields; disabled by default and not offered on the public instance

In development V0.3.2

Large-model command layer

A large-model command layer through an inference gateway, and ANet multi-agent delegation (moved out of the V0.3 seal)

Research D3.x / D4

3DGS appearance layer

Gaussian appearance layer from mesh distillation and real aerial assets; the G-3DGS evaluation gate decides product entry

Research D3.x / D4

Cross-entity collaborative experimentation

Heterogeneous entities (UAV, UGV, USV, ground robots, fixed sensors, software agents) in one experiment; declarative experiment specifications, replaceable collaboration algorithms and a collaboration leaderboard

Research D4

GPU reconstruction

Reconstruction engines such as LingBot-Map, DA3-Streaming and MapAnything (the mock engine pipeline already runs end to end)

Research V0.5

Real ANet network

A self-hosted ANetHub, one daemon per drone, CBBA allocation under disconnection

Research V1.0

Documentation map

You want to

Start here

Try it without installing

The public instance, with Shared show and Sandbox

Run it locally

Installation and requirements and Quick start

Control a sandbox from code

Python SDK and Open interface

Understand the internals

Architecture and coordinate frames and Worlds and data

Deploy your own instance

Deployment

Follow progress and quality

Versions and roadmap and Acceptance summary

The design documents are written in Chinese and live in the repository under docs/: the design baseline with 85 ADRs (docs/03), the D2 design supplement (docs/04), the D3 design supplement (docs/05), the module PRDs (docs/modules/) and the implementation and acceptance reports (docs/impl/). This site is written for users and summarizes the parts relevant to using the system; where the two differ, the design documents prevail.