Introducing Deft Robotics' unified deployment platform for physical AI. Frontier models are moving fast. Real world autonomy isn’t. Closing the gap takes more than better models. The missing piece? Infrastructure to deploy, intervene, learn from failures, and continuously improve. For the past year, we’ve been deploying in that gap... and building the stack to close it. Now, we’re opening it up. Hardware available now. Software in beta. Here’s how to use it for your use-case 🧵(1/7) ↓show more

Simba Gen 1: foundation model-native humanoid platform Humanoid hardware can’t be designed in isolation from robot foundation models. The hardware determines what the model can see, learn, and control, while the model reveals what the hardware actually needs—creating a tight feedback loop. Yet many humanoid platforms today are designed without that loop, leading to sensor configurations, mounting strategies and degrees of freedom that don't work in practice. That philosophy shaped Simba Gen1, our humanoid platform for manufacturing—tested over millions of cycles & validated through 500+ hours on real automotive production lines running state-of-the-art robot foundation models. It combines a holonomic base and lift, a 2-DoF active neck, flexible sensor and peripheral mounting options, hot-swappable batteries, expandable 12V/24V/48V power rails, and easily replaceable off-the-shelf components wherever possible. The result is a robust, modular and scalable platform designed to evolve alongside the models running it. See it in action ↓

2. Tether: The Transport Layer for Teleoperation Edge cases that improve autonomy only show up after you deploy. Human intervention closes the gap between model performance and deployment KPIs, enabling faster deployments and improvements on edge cases only real deployments expose. Today, robotics teams either build teleoperation themselves or buy into a closed stack. Tether takes a different approach: transport as a building block. It exposes low-latency, reliable transport for video, controls, and robot state through a simple API, abstracting the networking layer while keeping the teleoperator experience responsive as conditions change. Build whatever experience you want on top; run it on your infrastructure or ours. Tether powers Deft’s production deployments: 1,000+ intervention hours in the last year on real factory floors over cellular links, with robot–operator sessions spanning 5,500+ miles and 99.9%+ platform uptime. See it in action ↓

3. Eigen: the data flywheel for robot learning Models depend on high-quality data. But as deployments scale, maintaining that quality becomes increasingly difficult with human review alone. Eigen is the data platform that closes this loop, keeping quality high as volume grows without proportionally scaling human review. It ingests raw logs, automates rigorous quality checks, and produces curated datasets in popular open-source-compatible formats. At Deft, Eigen has processed thousands of episodes of robot data across diverse deployments with an average upload-to-dataset-ready time of 2 hours with only 5% of cases requiring human review. See it in Action ↓

4. Physical Intelligence X Deft: factory-proven foundation models Deft Robotics and Physical Intelligence are bringing general-purpose robot intelligence into real automotive factories. The task is line feeding: repeatedly picking up parts and placing them onto the production line in the correct position and orientation. It sounds simple. But parts shift around in the material bin as they’re picked. Each pick presents a different position and orientation, making fixed, rules-based automation impractical to deploy at-scale. Physical Intelligence’s robot foundation models, paired with Deft’s hardware and deployment stack, let the robot adapt to this variation. The robot locates each part, adjusts its grasp, and rotates it into position for placement—feeding the line without requiring every part to start in a predefined location. Across both deployments: 500 hours of operation and more than 10,000 parts processed. Every intervention surfaces the failures that matter, so the next model learns from real production failures. It's this iterative loop of deploying, intervening, retraining, and redeploying that makes the robot better over time — not just the cleanest demo. Robot intelligence becomes useful when it survives the factory.

5. Canary: edge-case detection and auto-annotation Real-world robot deployment is full of edge cases: dropped parts, failed grasps, placement errors. But capturing and labeling every one of these edge cases manually doesn’t scale. So how do you turn what happens on the robot into training data automatically? Canary watches every run in real time, knows exactly which subtask the robot is on, and alerts you when something goes wrong. Canary annotates subtasks automatically as they happen without a human annotator. So, successful runs come out as clean labeled demonstrations and every intervention comes out as a labeled failure case. Autonomously and continuously improve your robot workforce with Canary. See it in Action ↓

6. Colloid: intelligent observability layer for robot fleets Failures are inevitable in real robot deployments. Downtime is expensive. When something breaks, the answer is scattered across logs, metrics, alerts, and system state. The signals are there. Colloid connects them. It lives in your workflow like a teammate: monitors the fleet, cuts the noise, and surfaces what matters. One context layer for engineering, ops, support, and sales. Colloid already runs across Deft company-wide. See it in action ↓

Hardware available for purchase today. (2 week leadtime) Software available in beta. Sign up below ↓

congrats on the launch

Congrats @0xkharban on the launch!!

YES!!

Congratulations Deft Robotics team!

Epic

Let’s goooo

congrats on launch!

Can you check dm?

The real test will be what happens after deployment. Seeing how the system handles failures and learns from them will matter as much as the demo.

Congratulations 🎉 Go bears 🐻🥳

🚀🚀🚀🚀

congrats, great achievement shane!

@jungwonshin95 Congrats! Lots of cool stuff!

'intervene' is the load-bearing word. intervention data is the highest-signal and the most biased: every sample is conditioned on the policy already failing, and the label is an operator's recovery, not the right answer. the pre-intervention state matters as much as the fix.

Congrats Shane!

Big congrats on the launch, guys!

Physical ai finally getting the deployment layer it actually needs honestly ngl

@Scobleizer Awesomeness

Awesome!! I think frontier models moving fast while real world autonomy does not is the right framing I'd love your take on this, When an intervention fires, how do you tell whether it was the model, a calibration drift, or hardware wear? since retraining fixes the first and hides the other two. Congrats again on shipping!