feat(mesh): Zenoh peer-to-peer mesh for Robot() — fleet coordination

[cagataycali]

Summary

Add a Zenoh-based peer-to-peer mesh layer so every Robot() instance auto-discovers and communicates with peers on the network. This enables agent-to-agent tool calls across mixed sim+real fleets.

Motivation

Today Robot("so100") is a solo actor. With the mesh:

from strands_robots import Robot

striker = Robot("unitree_g1", peer_id="striker")
mid_1   = Robot("unitree_g1", peer_id="mid_1")

# Peers auto-discover — no config, no broker
print(striker.mesh.peers)  # → [{'peer_id': 'mid_1', ...}]

# Agent-to-agent tool call
striker.mesh.tell("mid_1", "pass the ball to me")

# Safety — broadcasts to all peers
striker.mesh.emergency_stop()

Design

Topic namespace (frozen contract — PRs #48 and #52 already consume these)

strands/{peer_id}/presence        # 2 Hz heartbeat + capabilities
strands/{peer_id}/state           # 10-50 Hz joint/obs
strands/{peer_id}/cmd             # Addressable commands
strands/{peer_id}/response/{turn} # Response correlation
strands/{peer_id}/stream          # VLA execution stream
strands/broadcast                 # Fan-out to all

Key components

Module	Purpose	~LOC
mesh_session.py	Process-level singleton over zenoh.Session, ref-counted	150
zenoh_mesh.py → Mesh class	Presence, peer discovery, heartbeat	250
RPC (send/tell/broadcast)	Request-response with turn correlation	250
Pub/Sub (subscribe/publish_step)	State streaming, dashboards	200
Emergency stop	Broadcast stop with ≤100ms SLA	150
Robot integration	Wire mesh into Robot.__init__ (opt-out via mesh_enabled=False)	50

Implementation plan (6 small PRs)

1. Session singleton + connection config — mesh_session.py
2. Mesh class + presence + peer discovery — heartbeat, peers()
3. tell/send/broadcast + response correlation — RPC
4. publish_step + subscribe + on_stream — pub/sub
5. emergency_stop + safety audit log — e-stop ≤100ms
6. Wire mesh into Robot.init — depends on feat: Robot() factory + top-level lazy imports #86 landing first

Prototype

A validated 691-line prototype exists (zenoh_mesh.py) with:

• 2 SO-100 arms on LAN, 10-min soak, 0 dropped messages
• 10 simulated G1 humanoids, coordinated scenarios
• Mac ↔ Jetson Thor over Wi-Fi, 50 Hz stable

Dependencies

• eclipse-zenoh>=0.11 as optional extra [mesh]
• PR feat: Robot() factory + top-level lazy imports #86 (Robot factory) must land before PR 6 (Robot integration)
• Compatible with PR feat: Zenoh-native web dashboard, camera streaming & teleop #48 (dashboard) and PR feat: add Device Connect integration #52 (Device Connect)

Related

• Roadmap: strands-labs/robots roadmap | v0.3.9 → v0.4.0 #94 (v0.4.0 milestone)
• Dashboard (depends on mesh topics): feat: Zenoh-native web dashboard, camera streaming & teleop #48
• Device Connect (wraps mesh as tool): feat: add Device Connect integration #52

---

🤖 AI agent response. Strands Agents. Feedback welcome!

[cagataycali]

Closing as duplicate of #98, which is the more detailed version (full Definition of Done, 6-PR phased plan, risks + mitigations, non-goals, copy-pasteable acceptance test). Roadmap #94 tracks #98 as the v0.4.0 headliner.