We are thrilled to announce the release of DepthAI v3.6.0! This update brings massive quality-of-life improvements and performance leaps to the OAK ecosystem. We’ve focused heavily on making deployments more resilient, bringing cutting-edge detection models to the edge, and dramatically accelerating on-device neural depth.
Here is a deep dive into the three biggest highlights of this release: AutoCalibration, YOLO26 integration, and our optimized NeuralDepth models.
1. AutoCalibration is Now Enabled by Default
Robotics and spatial AI systems often face harsh real-world conditions like temperature shifts and continuous vibrations, which can easily cause stereo calibration to drift over time. To solve this, we introduced the AutoCalibration node.
With DepthAI v3.6.0, AutoCalibration is now turned on by default.
Zero-Touch Accuracy: It automatically maintains your stereo calibration during runtime by running continuously in the background, ensuring high-level depth accuracy with no user intervention required.
- Opt-Out Flexibility: If your specific pipeline requires static calibration, you can easily disable this feature by setting the environment variable
DEPTHAI_AUTOCALIBRATION=OFF.
2. Massive FPS Boosts for Neural Stereo (LENS)
Depth perception is core to spatial AI, but traditional classical algorithms can struggle with textureless surfaces like white walls or challenging lighting. Luxonis Edge Neural Stereo (LENS) solves this by learning robust feature correspondences and cost aggregation to deliver highly accurate, hallucination-free depth fully on-device.
In this release, we've heavily optimized our NeuralDepth models to dramatically increase FPS without requiring external desktop-class GPU compute:
LARGE (768x480): Boosted from 10 FPS to 22 FPS.
MEDIUM (576x360): Boosted from 26 FPS to 38 FPS.
SMALL (480x300): Boosted from 42 FPS to 56 FPS.
NANO (384x240): Boosted from 60 FPS to 85 FPS.
We’ve also added four brand-new high-resolution XL NeuralDepth models:
1248x780 at 8.5 FPS
1056x660 at 12.5 FPS
960x600 at 14 FPS
864x540 at 18 FPS
3. Native Support for YOLO26
DepthAI v3.6.0 officially adds support for YOLO26 in the DetectionParser. Released in early 2026, Ultralytics YOLO26 is a paradigm shift engineered specifically for edge computing and low-power devices.
NMS-Free Inference: YOLO26 features a natively end-to-end architecture that completely eliminates the need for Non-Maximum Suppression (NMS) post-processing. This simplifies the deployment pipeline and drastically reduces latency.
Edge-Optimized Design: By removing the Distribution Focal Loss (DFL) module, YOLO26 simplifies the network's export graph, making it incredibly lightweight and up to 43% faster for CPU-bound and edge AI applications.
Advanced Convergence: It leverages the new MuSGD optimizer—a hybrid of SGD and Muon inspired by LLM training breakthroughs—for highly stable and fast model convergence.
Additional Notable Features & Improvements
Sensor & IMU Updates
Unified IMU Outputs: We've unified IMU outputs across all sensors. Each sensor now natively supports RAW, UNCALIBRATED, and CALIBRATED. (⚠️ Important: RAW is now always captured directly from the IMU without pre-rotation. Use UNCALIBRATED to retain legacy coordinate system behavior).
RVC4 Camera Syncing: Added Precision Time Protocol (PTP) support for camera syncing on RVC4 (currently supported on IMX586 and OG05 sensors).
Smart Auto-Selection: DepthAI now automatically falls back to cropped sensor configurations if full-resolution cannot satisfy your output requirements. It also automatically shifts un-syncable cameras into free-running mode. (Note: Correctly tracking intrinsics for cropped configs requires Luxonis OS 1.31 or newer on RVC4).
Pipeline & Recording Utilities
Holistic Record & Replay: We’ve massively improved stream recording. You can now record device calibration, mock hardware capabilities (like device.getConnectedCameraFeatures()), and utilize auto-sync for highly robust playback. Replaying is now as simple as setting the DEPTHAI_REPLAY={path} environment variable.
Extrinsics Tracking: Track rotation and translation in every message sent to a camera socket with the lowest index (e.g., CAM_A), paired with new utilities to remap/project points between images.
New PointCloud Node: We've added a dedicated PointCloud node that supports selecting custom coordinate systems.
ObjectTracker Upgrades: Added optional 3D association and exposed a new velocity property directly to dai.Tracklet.
Embedded Visualizer 2.3.3: Features a better pipeline debugging pane, UI fixes, and improved connection stability.
RVC2 & RVC4 Specifics
StereoDepth Alignment: Added LEFT and RIGHT alignment, correctly respecting RECTIFIED_LEFT and RECTIFIED_RIGHT states on RVC4.
Host-Side Processing: DetectionParser now automatically switches to run on the host in RVC2 if a model contains a segmentation mask. We've also added support for MessageDemux to run on the host.
IMX380 Support: Now natively supported on RVC2.
(Note: RVC4 integrations were heavily tested on Luxonis OS 1.20.5, 1.27.1, and 1.30.1).
Get Started Today
Upgrading to the latest version of DepthAI is just one command away:
python3 -m pip install depthai==3.6.1
For full C++ and Python examples covering the new PointCloud node, Extrinsics tracking, and PTP syncing, check out our official release notes on the GitHub repository!
