UniflexAI

Insight 9 SLAM Robustness Test 🎥 We put our AI-native spatial camera through rapid, dynamic movements. Watch the right screen: TinyNav VSLAM trajectory in RVIZ stays perfectly stable. 24G accelerometer & 188° ultra-wide FOV = zero tracking loss.#SLAM #Robotics #EmbodiedAI #VSLAM

FoundationStereo >> FastFoundationStereo > Retinify

Tuned a faster Harris frontend kernel on an A55 ARM CPU @ 1.5GHz. For 752×480 EuRoC frames: • 59.49 ms → 21.51 ms • 2.77x faster Harris Same output: • 159 detected points • 139 tracked points Pipeline runtime: • 68.43 ms → 43.20 ms

We expected FastFoundationStereo to be close to FoundationStereo. We were wrong. On Jetson Orin Nano: - FoundationStereo surprised us — lidar-quality point cloud - FastFoundationStereo is far worse (yes, it is indeed fast) - Retinify matches FastFoundationStereo's quality but is faster Speed: 100ms vs 15ms. Our fork fixed the 2-stage ONNX export with custom TensorRT plugins → single ONNX/plan. Repo: github.com/UniflexAI/Fast… cc @retinify @bowenwen_me

If you know your hardware well enough, you can still write code that's faster than today's AI. I wrote a Harris corner detector kernel that's 2× faster than OpenCV.

Just updated it to include frontend benchmarks. The first version compares Harris + Optical Flow with ORB + Hamming matching. Keep watching: github.com/dvorak0/slambe… You can also easily reproduce the test with: docker run --rm --cpuset-cpus="0" uniflexai/slambench:latest

Thanks for open-sourcing such great models. In our early test, Fast Foundation Stereo reached real-time performance (>13.5 FPS) with max disparity 64 on Jetson Orin NX — honestly much better than we expected. We’re planning to integrate it into our open-source navigation stack soon: github.com/UniflexAI/tiny…

I’ve updated github.com/dvorak0/slambe… with more benchmark results, including row-major vs column-major comparisons and GTSAM SmartFactor benchmarks. Added a summary screenshot below — check it out!

Working on SLAM for years, I’m constantly asked: which method is faster? So I made a one-line command to benchmark solver speed with the standard dubrovnik/problem-16-22106-pre.txt input: `docker run --rm uniflexai/slambench:latest` Run it, share your result, and tell me if anything looks unexpected. cc @gtsam4 @SkydioHQ @sandwichmaker

The wait is over. Meet Insight 9👁️ Native Spatial AI. 24g IMU. All packed in a Premium Aluminum Alloy Case. We are redefining vision for Embodied AI. Coming exclusively to Kickstarter. #SpatialAI #Robotics #Kickstarter #TechLeak #EmbodiedAI #TechNews #TechTrends

🚀 Introducing the TinyNav Bounty Program We’re rewarding the community for contributions that improve navigation, perception, and tooling in physical AI. 🧠 Bug reports 🧩 Code contributions 📝 Docs & tutorials 🎥 Demos & content Earn rewards while shaping the future of open robotics. 🔗 Details: docs.google.com/spreadsheets/d… #TinyNav #Robotics #OpenSource #PhysicalAI

Is it just me or are these robots getting TOO real? 🤖 Caught this @UnitreeRobotics dog zooming past. That new sensor rig looks like giant shades 🕶️. Scary smooth gait. Is this a secret field test? 🤯🔥 #Robotics #Unitree #TechLeak #FutureIsHere #Spotted

saw completely new design of the head. @UnitreeRobotics is going to navigate the robots using stereo camera, just like our open source project tinynav github.com/UniflexAI/tiny…

Excited to share a video about the workflow of our project! From zero to navigation with TinyNav 🗺️➡️🤖 1️⃣ Build a map 2️⃣ Add POIs in the editor 3️⃣ Let your robots navigate, correctly react to obstacles, and find feasible paths Keep an eye on TinyNav — a ~3,000-line open-source navigation stack for multiple robot platforms 🚀 We’ve added a remote interface using @foxglove — an amazing tool. Thanks, Foxglove 🙌 #Robot #Unitree #navigation

We need re-consider our stereo model! I will report the difference soon in our project.

@pablovelagomez1 @rerundotio @Gradio can't believe SAM works for so many classic vision problems.

Finished building out the @rerundotio and @Gradio app for SAM3D-Body, and I think it came out really clean! Under the hood, it's using three models 1. sam3d for exemplar segmentation based on the "person" text prompt. 2. sam3d-body for generating the 2d keypoints, 3d keypoints and mesh 3. mogev2 for intrinsic/fov estimation Really happy with how it came out. I'll probably start working on videos and multiview captures next!

Cool! New player in the room.