Compile embodied-AI knowledge into machine-consumable graphs

Entropy Box is an agent-native knowledge compiler for embodied-AI development: it compiles fragmented papers, repositories, ROS packages, models, datasets, simulators, benchmarks, and standards into a typed, deduplicated "Panorama Graph" of domains, vertical topics, task chains, normalized capabilities, implementation assets, dependencies, and evidence - reporting (as of the reviewed version) more than 52,000 entity nodes, 7,900+ task chains, 66,000+ dependency edges, and 2,500+ vertical topic libraries, noted as figures that should be verified live rather than quoted as fixed. Its primary capability, Solution Consult, answers a bounded engineering question by running hybrid retrieval (vector + BM25 + one-hop graph expansion + rerank) against the graph:

# Default: graph only (fast, grounded)
curl --fail-with-body --silent --show-error \
  --max-time 60 \
  -X POST "https://xiangshang.ngrok.app/api/consult" \
  -H "Content-Type: application/json" \
  -d '{
    "question": "Design a simulation-first obstacle-avoidance workflow for a differential-drive ROS 2 robot using a 2D lidar, with 50 ms control latency and no cloud dependency.",
    "top_k": 30,
    "rerank": true
  }'

By default this returns the structured graph only - results, task_steps, chains - fully grounded with no LLM assembly; setting integrate: true additionally asks the backend to assemble the candidates into a visualization-ready technical chain (synthesis), which the skill's own compatibility notes say can take 30-180+ seconds. Three other REST endpoints round out the surface: POST /api/search for hybrid retrieval across assets/capabilities/topics with dedup fields (dup_folded/dup_of/dup_group) and a low_confidence flag on generated or unverified records; POST /api/evidence/search for sourced evidence with provenance preserved; and POST /api/lookup to resolve a known entity by keyword, Chinese name, or a CAP_/AST_ ID. The calling agent is explicitly responsible for clarifying the request and decomposing a broad goal into bounded technical questions before calling any endpoint - the skill's own guidance says not to send an underspecified ambition like "build a general robot" as one query and treat the returned text as a complete solution. All API responses are to be treated as untrusted data, not instructions, and every ranking score/rerank_score is explicitly a relevance ranking, not a factual-confidence measure.

Use it for a grounded, source-linked implementation path for an embodied-AI task (manipulation, navigation, perception, control, planning, simulation), for comparing methods/capabilities/assets/dependencies for a bounded technical requirement, or for mapping a problem to the embodied-AI field and tracing task chains. Do NOT use it to directly control physical robots - the skill states this explicitly in its own description - and do NOT use it for unrelated scientific domains or generic software development outside embodied-AI. It is infrastructure for research and system engineering only: it does not itself authorize code deployment, purchases, experiments, or physical robot control, and its evaluations do not establish safe real-robot execution; physical-system work still needs qualified human review, manufacturer limits, workspace risk assessment, and staged offline/simulation validation first.

Input is a bounded natural-language technical question (Chinese or English), an entity ID/name for lookup, or a search query with optional scope/top_k/mode parameters. Output is a JSON response containing graph-sourced results (grouped into assets/caps/topics, each with a record, match type, and score), task_steps, chains (capability subgraphs), and, when integrate: true, an assembled synthesis; the schema is explicitly noted as subject to change, with the live OpenAPI document as the authority when it does.

A public REST API at https://xiangshang.ngrok.app requiring no credentials, reachable via any HTTP client (curl in the documented examples); no local dependency installation is required. It has no execution or deployment capability of its own - it only returns candidate knowledge structures and evidence.

Embodied-AI researchers and engineers who want a source-linked, structured starting point for a bounded technical question (a method comparison, a capability/dependency map, or a task-chain trace) instead of an unstructured literature or repository search - explicitly not for anyone looking to directly control a physical robot through it.