mirror of
https://github.com/ruvnet/RuView.git
synced 2026-10-18 21:00:22 -05:00
Add virtual-view prototype: integrations/virtual-view/README.md
This commit is contained in:
1 parent
f8750a4a50
commit
88daeb8cf2
1 file changed
+35
@@ -0,0 +1,35 @@
|
||||
# Virtual-view broker adapter (research prototype)
|
||||
|
||||
Opt-in integration with Dream Machine's `scripts/virtual-view-broker.mjs`.
|
||||
Disabled by default. No cameras, model calls, network, hardware commands or
|
||||
automatic promotion. It is NOT part of the Rust production perception pipeline.
|
||||
|
||||
```js
|
||||
import { generateKeyPairSync } from 'node:crypto';
|
||||
import { broker } from '../../../dream-machine/scripts/virtual-view-broker.mjs';
|
||||
import { createVirtualViewAdapter } from './adapter.mjs';
|
||||
// Example in-memory research identity only; use a trusted host signer in deployment.
|
||||
const { privateKey } = generateKeyPairSync('ed25519');
|
||||
const adapter = createVirtualViewAdapter({ broker, privateKey, enabled: true });
|
||||
// adapter.observe(calibratedSnapshot) returns RENDER, SKIP or BLOCK plus authority:none.
|
||||
```
|
||||
|
||||
The example assumes sibling RuView and Dream Machine clones; adjust the import
|
||||
from the executing module's location. Pass calibrated `virtual-view.v1` scene
|
||||
snapshots in room-local ENU metres. See Dream Machine ADR-0108 for every field,
|
||||
threshold, trust boundary, receipt and acceptance rule. The adapter does not
|
||||
transform raw `spatial.evidence.v1` RF Gaussians or `ruview.lidar.depth.v1` packets:
|
||||
neither supplies the independently tracked manipulation scene this method needs.
|
||||
|
||||
```sh
|
||||
node --test integrations/virtual-view/adapter.test.mjs
|
||||
# From the sibling Dream Machine checkout:
|
||||
node scripts/bench-virtual-view.mjs ../ruview/integrations/virtual-view/adapter.mjs
|
||||
```
|
||||
|
||||
The cross-repository benchmark exercises this adapter with the real broker,
|
||||
renders 256-object synthetic scenes, signs and replays receipts. It measures local
|
||||
CPU overhead only. Required robot evidence remains missing: +20 points on two of
|
||||
three held-out tasks, no regression beyond five points, added p95 below 150 ms,
|
||||
at least 95% pre-action drift detection, and replayable receipts for all trials.
|
||||
No receipt grants actuator authority. Do not wire this to live robot control.
|
||||
Reference in new issue
Block a user