Cortix AI

We hand intelligence to robots.

[Cortix AI @Cortixofficial]

Cortix AI is now part of NVIDIA Inception @NVIDIARobotics 🟢

We capture synchronized vision, EMG, and tactile data from real human work, so robots can learn how hands actually feel, not just how they look.

Building manipulation policies? DMs open.

#NVIDIAInception #PhysicalAI

[Cortix AI @Cortixofficial]

@trendforce Whichever hand design wins, someone still has to collect data for 22 DOF. Joint angles alone won't cover it, you need contact and force too.

[Cortix AI @Cortixofficial]

@IlirAliu_ Kind of humbling for everyone bolting more sensors on. Wonder how the blind policy handles a grasp that slips mid-insert.

[Cortix AI @Cortixofficial]

@eventualai @karanwxliaa Did they say what latency they get driving from Mexico? Curious at what point it starts messing up grasps.

[Cortix AI @Cortixofficial]

@praneethsamine 27 epochs surprised us too. On the GDP question, a lot of teams are betting on recording people at their actual jobs instead of teleop. Way cheaper per hour, but you have to recover the actions yourself.

[Cortix AI @Cortixofficial]

@vedantnair__ Curious how the sim from the sponge-wiping video got the contact force right. That's usually where video-to-sim drifts.

[Cortix AI @Cortixofficial]

@techniahqrobot Force feedback on the leader makes a big difference in demo quality. Does the felt load get saved with the demo too, or just joint positions?

[Cortix AI @Cortixofficial]

@ChongZzZhang @JayHe748646 @KehanWen170077 @leggedrobotics How much human data is "few" here? A handful of clips per skill, or more like an hour of mocap?

[Cortix AI @Cortixofficial]

@Jiahui_Yang6709 Liked your point about reading contact from commanded vs measured finger joints. Where does that stop working, thin or soft objects?

[Cortix AI @Cortixofficial]

@yuxuank02 One clip per skill is a really short leash. For throw-and-catch, does the clip need to get the release timing right, or does RL sort that out on its own?

[Cortix AI @Cortixofficial]

@ID_AA_Carmack Same problem gets worse once you put EMG and touch sensors next to the cameras. We've stopped trusting after-the-fact timestamp matching, a shared trigger at capture saves a lot of pain.

[Cortix AI @Cortixofficial]

@alex_kai2020 @angelaschoellig @BronarsToni @younghyo_park Links:
SpatialDex https://t.co/Dn0vGGbpwC
T-WLA https://t.co/V5KGLz6z72
Multi-view LAM https://t.co/4Cl74ecvMB
Blind assembly https://t.co/gBFMKUW1FC

[Cortix AI @Cortixofficial]

Papers we read this week on human data and contact for manipulation. Four of them, one number each.

[Cortix AI @Cortixofficial]

@alex_kai2020 @angelaschoellig Blind Precise Assembly from @BronarsToni and @younghyo_park: no cameras, no force or tactile sensors, only joint positions and velocities. Trained in sim, it inserts real pegs zero-shot in 73% of trials.

The joints already feel more contact than we usually give them credit for.

[Cortix AI @Cortixofficial]

We were at @roboflow's Visual Intelligence Summit today.

@josephofiowa framed the keynote as see, understand, act. The first two have gotten fast. They filmed a beach ball bouncing around the audience, autolabeled it, trained a model, and a few minutes later it was keeping score off a live drone feed. They also put out RF-DETR Pose for keypoints, Apache 2.0.

Act is where we work. A camera can tell you where every fingertip is. It can't tell you which frame the fingertip touched the part, or how hard the hand is squeezing. And once the hand goes behind the object it's guessing.

That's why we record sEMG, tactile and IMU next to the video.

[Cortix AI @Cortixofficial]

@wuji_global @nvidia Lots of labs will end up logging data from the same hand now. Does the Agile ONE build record fingertip touch too, and at what rate?

[Cortix AI @Cortixofficial]

@Robotiq_Inc Synced slip and force with labeled events is really useful. Any plans to export straight to LeRobot format?

[Cortix AI @Cortixofficial]

@LeoKharon What trained the action codec, robot teleop or retargeted human hands? Feels like that decides which grasps RL can even find.