The Three Layers

Every .flux holds three natures, each mapping to a standard format. Together, they make one DevReady asset — not a folder of drifting sidecars.

BODY — the physical structure

Holds the robot's physical description: links, joints, constraints, geometry, inertial properties, device-comm topology, and affordances.

kind	Description	Key meta
robot/link	A rigid body	name, mass
robot/joint	An edge (parent->child)	type (revolute|prismatic|fixed), lower, upper, axis
mesh	Geometry (binary STL)	— (BIN payload, ptype=model/stl)
device-comm/node	A bus or device	protocol (CAN|RS485|EtherCAT|MAVLink), baud, addr

Closed loops (4-bar, Delta, Stewart) are first-class: the graph model allows cycles, unlike URDF's pure tree. flux_robot_has_cycle() detects them.

Transcodes to: USD (flux_to_usd — meshes become def Mesh prims).

MIND — the knowledge + agency

Holds what the robot knows and can do:

kind	Description	Payload
concept	Task knowledge (OKF markdown)	markdown body
agent_card	An A2A agent card	JSON
task	A callable skill (linked part_of to its card)	JSON

An LLM/VLA reads MIND directly: flux_to_okf() produces a markdown bundle that an agent ingests for zero-shot task understanding. flux_to_a2a() produces an A2A agent-card.json for multi-agent orchestration.

Transcodes to: OKF + A2A.

JOURNAL — the lifetime log

An append-only log that grows for the robot's entire life:

kind	Description	Key meta
signal	A telemetry sample	name, value
param	A configuration parameter	name, value
phm_slice	A PHM (health management) time slice	timestamp range

Each record carries a multi-clock-domain clock field: sim_time / wall_time / device_monotonic — so sim and real-world telemetry can be aligned unambiguously.

The journal is what makes .flux a living digital twin: it grows on-device (MCU) as the robot operates. flux_compact can reclaim growth while preserving the live state.

Transcodes to: MCAP (rosbag2, primary) + MAVLink (edge transport; large assets automatically filtered off the bus).

Layer routing

Layers are a bitset on the layer field. A record can belong to one or more layers. Transcoders and scans filter by layer:

from fluxmeme import Store, LAYER_BODY, LAYER_MIND, LAYER_JOURNAL

with Store("robot.flux") as s, s.read() as txn:
    body_recs    = list(s.scan(txn, layer=LAYER_BODY))      # for USD
    mind_recs    = list(s.scan(txn, layer=LAYER_MIND))      # for OKF + A2A
    journal_recs = list(s.scan(txn, layer=LAYER_JOURNAL))   # for MCAP + MAVLink

---

三层模型(中文)

每个 .flux 承载三种本性,各映射到一个标准格式。三者合一构成 DevReady 资产—— 不是一个漂移的 sidecar 文件夹。

BODY — 物理结构

kind	说明	关键 meta
robot/link	刚体	name, mass
robot/joint	连接(父→子)	type, lower, upper, axis
mesh	几何(二进制 STL)	BIN payload
device-comm/node	总线/设备	protocol, baud, addr

闭链(四连杆/Delta/Stewart)是一等公民:graph 模型允许环,超越 URDF 纯树。 投影到:USD。

MIND — 知识 + 能动性

kind	说明
concept	任务知识(OKF markdown)
agent_card	A2A agent 卡(JSON)
task	可调用技能(链接 part_of 到卡)

LLM/VLA 直接读 MIND:flux_to_okf() 产出 markdown 包供 zero-shot 理解。 投影到:OKF + A2A。

JOURNAL — 终身日志

kind	说明
signal	遥测采样
param	配置参数
phm_slice	PHM(健康管理)时间切片

每条记录带多时钟域 clock:sim_time / wall_time / device_monotonic—— sim 与真实世界遥测可无歧义对齐。日志在设备(MCU)上随运维生长。 投影到:MCAP(rosbag2,主)+ MAVLink(边缘传输)。