Add NewModel/v2/deployment/RTC_INTERFACE_EN.md
Browse files
NewModel/v2/deployment/RTC_INTERFACE_EN.md
ADDED
|
@@ -0,0 +1,47 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# RTC deployment-interface verification
|
| 2 |
+
|
| 3 |
+
Date: 2026-09-18
|
| 4 |
+
|
| 5 |
+
## Result
|
| 6 |
+
|
| 7 |
+
The deployed msgpack/WebSocket interface supports RTC inference correctly for
|
| 8 |
+
both active action representations. Verification used the real policy server,
|
| 9 |
+
the public `PolicyClient`, real DiT-S ttRTC checkpoints on GPU, and localhost
|
| 10 |
+
wire serialization—not only direct Python wrapper calls.
|
| 11 |
+
|
| 12 |
+
| Representation | Tested policy | Required history | Delay 3 result |
|
| 13 |
+
| --- | --- | --- | --- |
|
| 14 |
+
| raw | FM ttRTC | physical `prev_action_chunk`, `[1,30,7]` | finite `[1,30,7]`; 3 fixed action rows |
|
| 15 |
+
| B-spline | joint-DD ttRTC | normalized `prev_control_rows`, `[1,18,7]` | finite `[1,30,7]`; 2 affected spans and 5 fixed cubic-support rows |
|
| 16 |
+
|
| 17 |
+
The server advertises the contract in its handshake. `PolicyClient` reads
|
| 18 |
+
`rtc_requires_previous_field` and sends the matching field through
|
| 19 |
+
`infer_realtime`. Delay is measured in raw 30 Hz action steps and is strictly
|
| 20 |
+
limited to 1–10. Base checkpoints reject realtime calls. The server keeps no
|
| 21 |
+
cross-client plan cache; the previous plan is explicit in every request.
|
| 22 |
+
|
| 23 |
+
## Important B-spline constraint
|
| 24 |
+
|
| 25 |
+
B-spline RTC must receive the normalized parameter rows returned by the prior
|
| 26 |
+
request. It intentionally rejects decoded 30×7 actions. Re-fitting and cubic
|
| 27 |
+
support expansion map raw delay 3 to two affected spans and five preserved
|
| 28 |
+
control rows, avoiding the incorrect assumption that raw action indices equal
|
| 29 |
+
spline-control indices.
|
| 30 |
+
|
| 31 |
+
## Hardening performed
|
| 32 |
+
|
| 33 |
+
The live wire test revealed that msgpack-decoded NumPy arrays can be read-only.
|
| 34 |
+
The deployment boundary now creates an owned float32 copy before converting
|
| 35 |
+
previous actions or control rows to torch tensors. This prevents transport
|
| 36 |
+
buffer aliasing and removes PyTorch's non-writable-array warning. Versioned
|
| 37 |
+
Hugging Face layouts now auto-resolve `vN/sidecars/{raw,bspline}`.
|
| 38 |
+
|
| 39 |
+
## Evidence
|
| 40 |
+
|
| 41 |
+
- `outputs/FULL_VISION_512/summary/deployment/rtc_websocket_validation.json`
|
| 42 |
+
- `outputs/FULL_VISION_512/summary/deployment_validation_dit_{s,b}.json`
|
| 43 |
+
- Full project test suite, including read-only transport arrays and versioned
|
| 44 |
+
release sidecar discovery.
|
| 45 |
+
|
| 46 |
+
This validates policy-server inference only; it does not constitute a
|
| 47 |
+
closed-loop robot safety or task-success validation.
|