# Rooms and state contracts

Experiments are the primary API product; rooms are resources that provide
compatible equipment, sensing and starting-state guarantees. Equipment access is
one part of the experiment alongside objects, instrumentation, starting state
and recorded outcomes. Physical simulation stages conditions on real hardware;
software simulation executes a numerical model. The following currently
qualified rooms all use software simulation.

Choose a room by its declared embodiment, clock, observations and starting state.
The API's current room/interface descriptor is authoritative for supported
commands and limits.

| Room | Current contract | Starting state and observations |
| --- | --- | --- |
| SO-100 session | `workshop.robot-live.run@5`; six channels, ten simulated seconds, up to 40 chunks and 120 resident wall seconds; physics pauses between actions. | Authored or exact private committed checkpoint; native cameras and measurements. |
| NexArm pair | `workshop.nexarm-live.run@1`; 12 channels across two arms; paused bounded command chunks. | Authored or exact private committed checkpoint; overview and two wrist cameras. |
| NexArm continuous | `workshop.nexarm-continuous.run@1`; wall-paced physics, bounded forecasts, latest-frame JPEG and optional VP8/WebRTC. | Authored initial state in the current lane; no durable continuation claim. Feedback generations gate reconnection. |
| Physical equipment | Planned commissioning and adapter qualification. | Exact kit, calibration, telemetry, watchdogs and local stops require hardware measurements. |

Earlier SO-100 `@1`–`@4` contracts remain available for compatibility. They have
different horizons/capture behavior; do not relabel a saved run as `@5`.
Installing dependencies does not enable any execution profile.

A nominal sensing target is not a delivered minimum. The paused NexArm
[customer-model exercise](results/CUSTOM_VLA_EXECUTION-2026-10-22.md) observed
about 1.3–1.5 source samples/second with larger gaps during commands, below the
fixture's strict 2 Hz requirement. Preserve this compatibility gap when selecting
capacity; successful model/control integration does not remove it.

## Room state is an explicit promise

| State contract | Meaning and present status |
| --- | --- |
| Authored start | Supported simulated initial state pinned by revision. |
| Exact private continuation | Supported qualified SO-100 and paused NexArm checkpoints for the same tenant and matching identity. |
| Variable start | SO-100 has bounded, saved seeded cube arrangements. It does not imply arbitrary physical starts. |
| Object presence / as observed | Product option: promise equipment and object inventory, then show actual state. Shared physical handover is unimplemented. |
| Self-reset or mechanical assist | Product options: a robot routine or room mechanism restores a declared condition. Each requires its own qualification. |
| Guaranteed exact physical start | Stronger optional room promise requiring measured reset/verification; not required for every future room. |

Room reset, fault recovery and turnover are separate operations. A reconnect
cannot silently reset, clear a fault or replace an uncertain command. Cross-tenant
state inheritance is not implemented.

Exact continuation advances from the current committed private head. It does
not establish replay or fork of an arbitrary earlier checkpoint for identical-start
A/B trials. Do not substitute sequential continuation for that stronger promise;
the [fresh-agent comparison](results/VLA_ROOM_SELECTION-2026-10-22.md) retained
this as an unresolved requirement.

## Planned physical experiment classes

These are operating contracts to implement and qualify, not changes to the
current simulated starts or checkpoint guarantees above.

The proposed facility layout places each robot setup in a separate contained
room, similar to a small cubicle or room with walls, an access door, task objects
and cameras inside. A paired-arm setup can share one experiment room. Shared
compute and service space may sit outside the rooms. The enclosure, access,
protection and recovery design still requires qualification for the actual
equipment; the concept does not establish a commissioned physical facility.

| Class | Intended promise | Admission and cost implications |
| --- | --- | --- |
| Best-effort / as observed | Declared installed configuration, reported inventory and current observations; no requested reset or exact object poses. | Trusted robot health/homing as applicable, protective systems, necessary views and a fresh suitability check remain mandatory. Restore minimal operability, reroute or reject an unusable scene under predeclared preflight/charge terms. |
| Initialized | One declared start reset with measured pose tolerances, calibration and initial snapshots. | Reserve operator/setup capacity; quote included versus additional resets. A two-hour minimum is a hypothesis to test against actual reset cost. |
| Managed campaign | Reserved repeated experiments with agreed fixtures, reset cadence, consumables and outputs. | Quote the complete campaign and resources; repetition does not imply exact physical determinism. |

Separate installed configuration from inventory and current pose/scene state.
Object presence alone establishes neither visibility nor reachability. The
planned scene record should identify configuration, observation time, camera
coverage, robot health/homing, inventory and uncertainty, suitability assessment,
initialization class and any promised tolerances. Record who or what supplied an
assessment; unavailable confidence or calibration remains unknown. Do not turn
model estimates into measured facts or reuse a stale preview after intervening
work. Cross-customer scene reuse needs explicit privacy/cleanup handling.

Use existing scheduling authority to reserve physical equipment, trusted local
control, compatible GPU/CPU and required setup/reset windows without double
counting shared capacity. Failed preflight must not silently become charged
execution. Best-effort pricing, reset charges and cancellation/refund treatment
need disclosed terms and verification before activation. A short-run US$5 minimum
is a hypothesis, not an implemented tariff. Independent local protection applies
to every class and price.

Commercial room SKUs may combine robot, gripper, cameras, fixtures and a declared
state contract. Customers buy an experiment against those capabilities. Optional
onsite policy compute is a later paid capability.
[Pricing preview](ROOM_PRICING.md) and [roadmap](ROADMAP.md).

## Discovery and SDK direction

Published configurations now expose actual robot/controls, named RGB cameras,
timing/calibration/state promises, adapter requirements, missing features
and version-matched docs through [HTTP/SDK/MCP](ROOM_API.md#room-capabilities-and-independent-observations).
A customer agent can assess a fixed experiment
against several candidates directly; an optional concierge can help using the
same information. Fit, qualification and available capacity are separate checks.

[ADR 0057](architecture/0057-room-capabilities-and-independent-streaming-sdk.md)
defines clients independent of these arms, with individually admitted robot and
model adapters. SDK 0.1.6 `ObservationClient` reads original named RGB cameras
without arm-control imports or motion authority across SO100 capture/session and
paused/continuous NexArm. `RobotClient` remains specific to continuous NexArm.
Generic video negotiation, new sensors and source-bound device-inference results
remain planned; current execution/state contracts remain authoritative.