The Humanoid Interop Standard (HIS) — draft

Layer 0 live · Layer 3 draft + mock · Layers 1,2,4 unstarted

A proposed open standard for describing humanoid robot capabilities and letting LLM/VLA controllers discover and command any compliant robot without vendor-specific glue code — the same way MCP lets one LLM host talk to any tool server. This is a draft for discussion, not a ratified spec. Layer 0 — the capability manifest — is live, generated from the same cited data the XHI ranking already scores. Layer 3 — the tool-calling contract — has a draft schema and a working reference implementation in the site's repository, but is not connected to any real robot; no robot in this index implements a control protocol yet. Nothing below competes with ROS 2 or MCP; it's meant to be a thin certification/orchestration layer on top of both.

Layering

Layer	Purpose	Status
0	Embodiment taxonomy — what a robot physically is	live
1	Perception/state interface — how a robot reports sensing & status	not started
2	Action interface — how a controller commands a body (capability-negotiated, not identical schemas)	not started
3	Orchestration/tool contract — MCP-shaped tool discovery for LLM control, layered on ROS 2 for real-time transport	draft + mock reference
4	Multi-agent/fleet coordination — cross-vendor robot-to-robot coordination	not started
5	Certification — compliance levels, conformance testing, badging	schema only

Layer 0: the capability manifest

Every robot in the index already exposes a live, generated manifest — a machine-readable "spec sheet" an orchestrating LLM could read before issuing any control-protocol call. It's derived directly from the same fields the XHI ranking scores, not a second dataset to maintain.

Schema: robot-manifest.schema.json (JSON Schema, draft 2020-12)
Live for any robot: GET /robot/{slug}/manifest.json
Static examples: figure-02, tesla-optimus, softbank-pepper, agibot-d2-max

Form-factor distribution, this index

Classified per-robot from each robot's own cited sourcing — not inferred. One robot (AgiBot D2 Max) is correctly tagged quadruped even though it's filed in a humanoid ranking, because that's what its own sources say it is.

form_factor	Robots
bipedal	109 / 175
wheeled-base	26 / 175
other	13 / 175
quadruped	8 / 175
arm-only	5 / 175
differential-drive	4 / 175
wheeled-biped	3 / 175
usv	2 / 175
wheeled-heavy	2 / 175
articulated-arm	1 / 175
multirotor	1 / 175
multi-arm-cart	1 / 175

Try it — live manifests across the diversity of the dataset

Figure 02 · bipedal

/robot/figure-02/manifest.json

Optimus (Gen 2/3) · bipedal

/robot/tesla-optimus/manifest.json

Pepper · wheeled-base

/robot/softbank-pepper/manifest.json

AgiBot D2 Max · quadruped

/robot/agibot-d2-max/manifest.json

Layer 3: the tool-calling contract (draft, not connected to a real robot)

Four JSON-RPC methods — initialize, tools/list, tools/call, tools/confirm — modeled on MCP's tool-discovery shape, plus a risk_tier on every tool that a generic tool protocol doesn't need but a robot control protocol can't skip: a bad MCP call corrupts a file, a bad robot call can move mass near a human. A call whose risk tier is at or above the robot's declared confirmation_required_above_tier returns pending_confirmation instead of executing — the orchestrator must make a second, explicit call to actually run it.

The reference implementation is a simulated Figure 02 tool server exposing four Tier-1 skills (speak, navigate_to, grasp, hand_off) with real gating behavior — every call is asserted end-to-end, not just documented. It lives in the repository, not on this server, because there's nothing real for it to control yet.

Spec: standard/layer3-tool-contract.md
Tool schema: tool-definition.schema.json
Run it: python -m standard.reference.demo — prints every JSON-RPC message in the full initialize → tools/list → tools/call → tools/confirm flow

Certification levels (Layer 5 — schema exists, no test suite yet)

Level	Meaning
`documented`	Self-reported manifest exists, cross-checked against ≥2 sources like every other XHI figure — no live protocol test performed.
`conformant`	Passes an automated conformance suite: tool discovery resolves, schemas validate, safety-tier confirmation gating actually blocks correctly.
`interop-proven`	A reference orchestrating LLM that has never seen this robot before completes a benchmark task suite zero-shot — true demonstrated interoperability. The suite this level was written against, and never able to name, now exists in draft: NIST's Humanoid Robot Baseline Performance Benchmark, whose four capability areas include exactly the reasoning and scene-understanding axis a zero-shot orchestration claim turns on.

Every robot on this site is currently documented — no Layer 3 control protocol exists yet for any of them, so conformant and interop-proven are aspirational levels defined in the schema, not achieved by anything today. interop-proven stays aspirational for a second, independent reason: NIST has published the benchmark's four capability areas but not its task list or metrics, which are stated for release in summer 2026. See Benchmarks for what is and is not published.

This is a draft. Layers 1, 2, and 4 are not designed yet, and Layer 3's reference implementation is a simulation, not a connection to real hardware. If you're a robot vendor, an agent-framework builder, or a researcher and want to shape this, the schemas, generator, and reference tool server are all open in the site's repository under standard/ — feedback welcome via the same sourcing/citation standard the rest of this site holds itself to.