# Robot AI component contract

Original blank worksheet. The housing example is fictional; no system or performance result is demonstrated here.

## Application boundary
- Approved task and workpiece variants:
- Site, hardware, software versions, and review date:
- Conditions requiring human intervention:
- Application owner and independent machine-safety assessment owner:

## Complete one record per component
| Field | Enter evidence, not a marketing category |
|---|---|
| Component role | Action policy / world generation / embodied reasoning / controller / inspection / other |
| Exact product and version | Include availability and contractual support |
| Input | Format, source, freshness, missing-input behavior; robot state/embodiment where required |
| Output | Text, proposed tool request, numerical action schema, generated media, or measured result |
| Authority | What may consume the output? Who validates or rejects it? |
| Access and location | API / downloadable weights / trusted-tester access; device, site server or cloud |
| Runtime | Serving integration, inference hardware/memory, precision, OS; distinguish training hardware |
| Action consumer | Embodiment, dimensions, units, coordinates, update timing and supported controller |
| Timing | Observation timestamp and age at execution; inference time; actions returned versus executed per request; action execution rate; blocking or overlapping replanning |
| Failure behavior | Unsupported input, timeout, uncertainty, interrupted request |
| Evidence | Observed / vendor-reported / unknown; source and date |
| Maintenance | Update owner, version rollback process, retest trigger |

## Fictional example
Task: place an approved housing into an inspection tray.
- Reasoning: goal + observations → proposed approved routine, no direct motion authority.
- Prediction, if used: candidate action + state → estimated outcome, not measured success.
- Action policy: observations + task → proposed action representation.
- Controller/application: validated request + measured state → execution or rejection within assessed limits.

These are separate role contracts, not a required processing sequence. Do not use this worksheet as an operating instruction.

## Review decision
- Observed completion evidence:
- Unresolved supplier questions:
- What will be measured in the pilot:
- Person approving the next evaluation step:

## Documentation-based shortlist — reviewed September 6, 2026
- Gemini Robotics 2: vision/language to motor control; early-access partners. Confirm exact interface and deployment terms.
- Gemini Robotics ER 2: multimodal input to text; API/AI Studio route; Enterprise Agent Platform private preview.
- Gemini Robotics On-Device 2: text/images/numerical robot state to numerical actions; local inference; trusted testers. Public card does not supply a general inference hardware bill of materials.
- Cosmos 3 Reasoner: text/vision to text. Generator: supported visual/sound/action generation. Choose a specific supported integration and model variant.
- Cosmos3-Super 64B, Nano 16B and Edge 4B have different hardware recommendations. Consult the matrix; model size alone does not establish runtime requirements.
- Policy-DROID variants require a compatible embodiment/interface; a JSON action array is not a controller-ready instruction by itself.
Sources:
https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/
https://deepmind.google/models/model-cards/gemini-robotics-er-2/
https://deepmind.google/models/model-cards/gemini-robotics-on-device-2/
https://docs.nvidia.com/cosmos/latest/cosmos3/model_reference.html
https://docs.nvidia.com/cosmos/latest/cosmos3/model_matrix.html

## Timing example — vendor-reported, not a local test
NVIDIA's August 19, 2026 Cosmos 3 Edge tutorial reports approximately 1.53 seconds to generate a 32-action chunk on Jetson AGX Thor T5000. At 15 actions per second, the entire chunk spans approximately 2.13 seconds. The action execution rate is not the model inference frequency.
Record the executed prefix and replanning schedule separately. The tutorial's example executes only 16 actions before requesting another chunk; do not infer continuous motion merely from the full 32-action duration. Measure the end-to-end controller behavior, including whether inference overlaps execution.
Source: https://developer.nvidia.com/blog/post-train-nvidia-cosmos-3-edge-for-on-device-robot-control/

## Fictional interface record — not executable vendor JSON
Observation: tray-017.jpg and current station-state timestamp.
Reasoning output: left compartment appears empty; right compartment obscured.
Application decision: request a new observation; no placement command yet.
Optional policy output: numerical sequence whose schema, units, frame, timing and embodiment must match the consumer.
Completion: independently observed inspection result, never a predicted frame.