Repository navigation
Fix Gemini 2 depth stream assembly on Forge 1 / Dora 0.4.1 - #4
Conversation
mrgh97
left a comment
There was a problem hiding this comment.
Technical review for Forge 1 / Dora 0.4.1 and v1.1.1. No blocking findings. I checked capture readiness and recovery, the native Gemini 2 bulk Y16 bounds checks and diagnostic counters, the two-frame publication queue, and the embedded SDK build and license notices.
The original USB 3.0 Gemini 2 on firmware 1.4.98 passed 100/100 starts with zero original SDK errors. The previous functional head's packaged dataflow delivered 29.920 fps for 74.86 s and exited cleanly. The node reported the USB disconnect in about one second during a firmware reboot. The version changes preserve the locked runtime dependencies. Local tests passed (115 tests and 81 subtests). The merged-tag package will be validated before publication as a Forge 1 stable release.
Fixes the malformed Gemini 2 bulk Y16 depth assembly tracked in #2. This PR targets Forge 1 / Dora 0.4.1 on
stable/1.x. See #3 for the other Forge branch.The USB trace on the reporter's camera showed depth bulk fragments without a valid UVC header. LibUVC read those fragments as headers and forwarded short or oversized frames to the SDK. Linux package builds now compile the native SDK 2.10.6 from a pinned source revision with Gemini 2 bulk Y16 header, frame-boundary, and exact-size checks. Rejected headers and frames remain visible in the SDK shutdown summary. The binary includes the SDK and its license notices. The Python dependency remains
pyorbbecsdk2==2.1.3.The camera node also retains two encoded RGBD sets to absorb tick jitter. It publishes one set per tick and drops the oldest set if the two slots fill. For an explicitly selected USB serial, the isolated backend reports loss of the original USB address and stops the node.
Hardware regression
Tested the original Gemini 2, serial
AY6Z363011Y, firmware 1.4.98, on the same VIA USB 3.0 controller at 5 Gbps. The original 1280×720 RGB at 30 fps, 640×400 Y16 at 30 fps, hardware D2C and frame sync configuration was used throughout.STREAMINGat 0.215 s.The re-enumeration test used a firmware reboot with the USB cable in place. The 100-start run used the SDK patch with Python wheel bindings, while the continuous flow used the standalone Linux package.