masterleopold commited on
Commit
1d56382
·
verified ·
1 Parent(s): b239a93

v8 card: add candidate-path closed-loop SITL results (documentation only)

Browse files
Files changed (2) hide show
  1. README.md +17 -3
  2. release-manifest.json +2 -2
README.md CHANGED
@@ -130,6 +130,20 @@ that post-hoc rerun.
130
  These are synthetic scenes built to match the training renders, not real-flight
131
  evidence. See `evidence/multi-scene-sitl.json`.
132
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
133
  ## Loading
134
 
135
  Use the pinned Cosmos Framework revision. It registers the native
@@ -183,9 +197,9 @@ labels, and raw training data are not distributed in this model repository.
183
 
184
  - The final result is synthetic and covers two mission profiles.
185
  - There is no independent real-image final evaluation for this checkpoint.
186
- - SITL covers 8 synthetic scenes of one layout, one completed only after a post-hoc
187
- flight rerun. The flight adapter cannot yet fly programs that compile to more than one viewpoint, which
188
- a boundary observation of an occluded target needs.
189
  - Jetson Orin NX latency, memory, power, and concurrent-load behavior remain unmeasured.
190
  - Exact routes depend on trusted geometry, policy, vehicle limits, and the deterministic
191
  compiler; the model alone cannot produce an approved flight plan.
 
130
  These are synthetic scenes built to match the training renders, not real-flight
131
  evidence. See `evidence/multi-scene-sitl.json`.
132
 
133
+ On 2026-09-30 the same checkpoint was flown through the **candidate-selection
134
+ path**: its candidate readout chose between two supplied observation programs
135
+ instead of the Generator writing one. With the viewpoint compiler now aiming
136
+ whole-boundary observations at the target centre (an opt-in scene contract),
137
+ four preregistered procedural slope scenes (seeds 424-427) all reached a
138
+ completed decision with settings frozen; the one-sided 95% lower bound on the
139
+ success rate is about 47%. In a procedural road scene, a candidate readout head
140
+ fitted on synthetic features chose the close two-viewpoint program, and both
141
+ programs flew with every geometry check agreeing. These runs used procedural
142
+ scenes; the slope candidates compile to the same viewpoint, and the road head is
143
+ weak and does not reject. See the
144
+ [Decision Preview card](https://huggingface.co/v13s/tancho/tree/main/decision-preview)
145
+ for the run table.
146
+
147
  ## Loading
148
 
149
  Use the pinned Cosmos Framework revision. It registers the native
 
197
 
198
  - The final result is synthetic and covers two mission profiles.
199
  - There is no independent real-image final evaluation for this checkpoint.
200
+ - Generator-path SITL covers 8 synthetic scenes of one layout, one completed only after a
201
+ post-hoc flight rerun. Programs that compile to more than one viewpoint have since flown in
202
+ the candidate-path runs; their capture check was corrected during those runs.
203
  - Jetson Orin NX latency, memory, power, and concurrent-load behavior remain unmeasured.
204
  - Exact routes depend on trusted geometry, policy, vehicle limits, and the deterministic
205
  compiler; the model alone cannot produce an approved flight plan.
release-manifest.json CHANGED
@@ -34,8 +34,8 @@
34
  },
35
  {
36
  "path": "README.md",
37
- "bytes": 9871,
38
- "sha256": "08e35744364ef6ceea819ae8ba459c2290be4f5260dca62e366c5b8b59cca8c2"
39
  },
40
  {
41
  "path": "REPRODUCE.md",
 
34
  },
35
  {
36
  "path": "README.md",
37
+ "bytes": 10808,
38
+ "sha256": "667d04055ab8adc8175b9b66648a8df5e3b4acfe500c9ef46534021c73cbb153"
39
  },
40
  {
41
  "path": "REPRODUCE.md",