Polaroid

Embeddable CRDT scene graph for embodied AI agents.

CI PyPI version Python 3.10+ Downloads License: MIT codecov Typed

Quick Start · How It Works · CLI Reference · GitHub Action · vs. Alternatives · Contributing

---

Why

Multiple robots navigating the same building each build their own private map. When robot A opens a door and robot B hasn't been told, they diverge. Sharing a map requires a central server — which is a single point of failure.

polaroid solves this with a CRDT scene graph: a persistent, mergeable map of nodes (objects, rooms, surfaces) and edges (spatial relationships). Two robots can merge their maps without a server, without conflicts, without data loss. CRDT semantics guarantee the merge is always safe, deterministic, and idempotent.

# Share your scene graph with a peer
polaroid merge /path/to/peer/scene.db

---

How It Works

flowchart LR
    A[Agent observes\nroom / object / surface] --> B[SceneNode added\nto SceneStore]
    A --> C[SceneEdge added\ncontains / adjacent-to]
    B & C --> D{Peer agent\nhas different view}
    D --> E[SceneMerger.merge\nCRDT semantics]
    E --> F[Grow-only nodes\nconfidence-weighted LWW]
    F --> G[Unified scene graph\nno server required]

Core primitives:

• SceneNode — a content-addressed node (object, room, surface, region, or agent). ID = SHA-256[:16] of label|node_type. Same label and type always produce the same ID regardless of agent.
• SceneEdge — a directed spatial relationship between two nodes (contains, adjacent-to, on-top-of, blocks, connects). ID = SHA-256[:16] of source_id|target_id|relation.
• SceneStore — SQLite-backed persistent store. Zero dependencies beyond Python stdlib + click/rich.
• SceneMerger — CRDT merge: nodes are grow-only (never deleted), conflicting property updates resolved by confidence-weighted last-write-wins.
• SceneQuery — query by type, label substring, confidence, or spatial neighbors.

---

Features

Feature	Details
Content-addressed nodes	Same label+type always produces the same ID — no duplicates
CRDT merge semantics	Grow-only sets + confidence-weighted LWW registers
Conflict-free merge	merge() is idempotent, commutative, and associative
Spatial queries	Find nodes by type, label, or neighbors via edge traversal
Context summary	One-call text description of the scene for LLM prompts
Offline / local-first	Single SQLite file, no server required
FastAPI REST server	/node, /edge, /nodes, /merge, /context endpoints
MCP server	Model Context Protocol integration for Claude and other agents
202 tests	Comprehensive test suite covering all layers

---

Quick Start

pip install polaroid-ai

from polaroid import SceneNode, SceneEdge, SceneMerger, SceneQuery, SceneStore

# Robot A observes a kitchen
store_a = SceneStore("/tmp/robot-a.db")
kitchen = SceneNode(label="room-kitchen", node_type="room", properties={"floor": "tile"})
table = SceneNode(label="table-A", node_type="object", properties={"color": "brown"}, confidence=0.9)
store_a.upsert_node(kitchen)
store_a.upsert_node(table)

edge = SceneEdge(source_id=kitchen.id, target_id=table.id, relation="contains")
store_a.upsert_edge(edge)

# Robot B observes the same room with a door
store_b = SceneStore("/tmp/robot-b.db")
store_b.upsert_node(kitchen)  # same ID — no duplicate
door = SceneNode(label="door-1", node_type="object", properties={"state": "open"})
store_b.upsert_node(door)

# Merge B into A — CRDT guarantees safety
result = SceneMerger().merge(store_a, store_b)
print(result.summary())
# Added 1 nodes, updated 0 nodes, added 0 edges, resolved 0 conflict(s).

# Query the unified scene
q = SceneQuery(store_a)
print(q.context_summary())
# 1 rooms, 2 objects. Known objects: table-A, door-1. 1 spatial relationship recorded.

store_a.close()
store_b.close()

---

CLI Reference

polaroid [--db PATH] COMMAND [OPTIONS]

Command	Description	Key options
add-node LABEL TYPE	Add a node to the scene	--confidence FLOAT, --property K=V, --agent-id STR
add-edge SOURCE TARGET RELATION	Add a directed edge	--confidence FLOAT
query	Query nodes	--type TYPE, --label LABEL, --min-confidence F, --format {rich,json}
merge OTHER_DB	Merge another scene store into this one	—
status	Show node/edge counts and context	—

Global options:

Option	Default	Env var
--db PATH	.polaroid/scene.db	POLAROID_DB

Examples:

# Add nodes
polaroid add-node door-1 object --confidence 0.95 --property state=open --property color=brown
polaroid add-node room-kitchen room

# Add an edge
polaroid add-edge <door-id> <kitchen-id> contains

# Query the scene
polaroid query --type object
polaroid query --label door --min-confidence 0.8 --format json

# Merge peer's scene
polaroid merge /path/to/peer.db

# Status overview
polaroid status

---

GitHub Action

Add polaroid scene merge to your CI pipeline:

# .github/workflows/polaroid.yml
name: polaroid scene check
on: [push, pull_request]

jobs:
  scene:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: sandeep-alluru/polaroid@main
        with:
          db: .polaroid/scene.db
          fail-on-empty: "false"

The action installs polaroid and runs polaroid status. See docs/github-action.md for full documentation.

---

vs. Alternatives

	polaroid	ROS 2 map server	Semantic Fusion	Hydra (Facebook)	LangGraph checkpointing
CRDT merge	Yes — grow-only + confidence LWW	No	No	No	No
Serverless	Yes — single SQLite file	Requires ROS master	Requires GPU	Requires server	Partial
Content-addressed IDs	Yes — SHA-256[:16]	No	No	No	No
MCP / LLM integration	Yes — MCP server	No	No	No	No
Offline / embedded	Yes	Partial	No	No	Partial
Primary purpose	CRDT scene graph for multi-agent	ROS navigation maps	Dense 3D fusion	Neural scene representation	LLM state persistence
Open source	MIT	Apache 2.0	Research	BSD	Apache 2.0

polaroid is not a 3D reconstruction system. It is designed for: "Given that multiple agents observed different parts of the world, how do we merge their maps safely?"

---

Claude / MCP integration

polaroid ships a Model Context Protocol server that lets Claude and other MCP-compatible agents record and query scene nodes directly:

# Start the MCP server
python -m polaroid.mcp_server

# In your Claude Code project's .claude/settings.json:
{
  "mcpServers": {
    "polaroid": {
      "command": "python",
      "args": ["-m", "polaroid.mcp_server"]
    }
  }
}

Once connected, Claude can call add_scene_node, query_nodes, and get_context as tools. See docs/mcp.md for the full tool schema.

---

OpenAI integration

polaroid exposes a FastAPI REST server compatible with OpenAI's function-calling format. The tool definitions are in tools/openai-tools.json and the full API spec is in openapi.yaml.

# Start the REST server
uvicorn polaroid.api:app --reload

# Pass to Codex CLI or any OpenAI-compatible agent
codex --tools tools/openai-tools.json "Show me all objects in the scene"

Endpoints: GET /health, POST /node, POST /edge, GET /nodes, POST /merge, GET /context. See docs/openai.md for details.

---

Case Studies

See how teams are using polaroid in production:

• Conflict-Free Fleet Mapping for 60 Autonomous Warehouse Robots
• Persistent NPC World State for an AI-Driven Open-World RPG

---

Repository structure

polaroid/
├── src/
│   └── polaroid/
│       ├── graph.py          # SceneNode, SceneEdge, MergeResult dataclasses
│       ├── store.py          # SQLite-backed SceneStore
│       ├── merger.py         # SceneMerger CRDT merge algorithm
│       ├── query.py          # SceneQuery — find_nodes, find_neighbors, context_summary
│       ├── export.py         # to_dot(), to_json(), to_adjacency_matrix() exporters
│       ├── stats.py          # GraphStats, compute_stats(), cluster_by_type(), most_connected()
│       ├── subgraph.py       # extract_subgraph(), filter_by_type(), neighborhood()
│       ├── report.py         # print_scene(), print_merge(), to_json(), to_markdown()
│       ├── cli.py            # Click CLI (add-node, add-edge, query, merge, status, stats, export)
│       ├── api.py            # FastAPI REST server
│       └── mcp_server.py     # MCP server
├── tests/
│   ├── test_graph.py         # SceneNode, SceneEdge, MergeResult unit tests
│   ├── test_store.py         # SceneStore upsert/get/list tests
│   ├── test_merger.py        # SceneMerger CRDT merge tests
│   ├── test_query.py         # SceneQuery tests
│   ├── test_export.py        # Export formatter tests
│   ├── test_stats.py         # Graph analytics tests
│   ├── test_subgraph.py      # Subgraph extraction tests
│   ├── test_report.py        # Report formatter tests
│   ├── test_cli_runner.py    # Click CliRunner tests
│   └── test_api.py           # FastAPI TestClient tests
├── examples/
│   └── demo.py               # Standalone demo script
├── docs/                     # MkDocs documentation
├── tools/
│   └── openai-tools.json     # OpenAI function-calling tool definitions
├── assets/
│   ├── hero.png              # README hero image
│   └── logo.png              # Project logo
├── action.yml                # GitHub Action
├── openapi.yaml              # OpenAPI 3.1 spec
├── pyproject.toml            # Package metadata + dependencies
└── CONTRIBUTING.md           # Contribution guide

---

Advanced API

These functions are exported at the top level (from polaroid import ...) and cover graph analytics, DOT export, and subgraph extraction.

compute_stats(store) -> GraphStats

Returns aggregate statistics about a SceneStore.

from polaroid import SceneStore, compute_stats

store = SceneStore("/tmp/scene.db")
stats = compute_stats(store)
print(stats.node_count)          # total nodes
print(stats.edge_count)          # total edges
print(stats.avg_confidence)      # mean confidence across all nodes
print(stats.most_common_type)    # node type with the highest count

to_dot(store) -> str

Serialises the scene graph as a Graphviz DOT string, ready for rendering with dot -Tpng.

from polaroid import SceneStore, to_dot

store = SceneStore("/tmp/scene.db")
dot_src = to_dot(store)
print(dot_src)
# digraph polaroid {
#   "abc123" [label="kitchen (room)"];
#   "def456" [label="table-A (object)"];
#   "abc123" -> "def456" [label="contains"];
# }

with open("scene.dot", "w") as f:
    f.write(dot_src)
# Then: dot -Tpng scene.dot -o scene.png

extract_subgraph(store, node_ids) -> SceneStore (in-memory)

Returns a new in-memory SceneStore containing only the specified nodes and the edges that connect them.

from polaroid import SceneStore, SceneNode, extract_subgraph

store = SceneStore("/tmp/scene.db")
# Get IDs of interest from a query, then extract
kitchen = SceneNode(label="room-kitchen", node_type="room")
table   = SceneNode(label="table-A",      node_type="object")
sub = extract_subgraph(store, [kitchen.id, table.id])
print(sub.list_nodes())   # only kitchen + table

neighborhood(store, node_id, radius=1) -> list[SceneNode]

Returns all nodes reachable from node_id within radius hops (BFS over edges). Useful for building local context windows for LLM prompts.

from polaroid import SceneStore, SceneNode, neighborhood

store = SceneStore("/tmp/scene.db")
kitchen = SceneNode(label="room-kitchen", node_type="room")
nearby = neighborhood(store, kitchen.id, radius=2)
for node in nearby:
    print(node.label, node.node_type)

---

GitHub Topics

Suggested topics for discoverability:

ai-agents crdt scene-graph spatial-memory robotics embodied-ai sqlite mcp openai llm-tools multi-agent python

---

Star History Chart

---

Stay Updated

Subscribe to The Silence Layer — weekly dispatches on production AI infrastructure, new releases, and the failure modes that production AI systems don't surface until it's too late.