AI Takes a Bow: Interactive GLaDOS Robot Among 9 Winners in Hackster.io Challenge

YouTube robotics influencer Dave Niewinski has developed robots for everything from driveable La-Z-Boy chairs to an AI-guided cornhole tosser and horse-drawn chariot racing.

His recent Interactive Animatronic GLaDOS project was among nine winners in the Hackster AI Innovation Challenge. About 100 contestants vied for prizes from NVIDIA and Sparkfun by creating open-source projects to advance the use of AI in edge computing, robotics and IoT.

Niewinski won first place in the generative AI applications category for his innovative robot based on the GLaDOS guide from game series Portal, the first-person puzzle platform from video game developer Valve.

Other top winners included contestants Andrei Ciobanu and Allen Tao, who took first prize in the generative AI models for the edge and AI at the edge applications categories, respectively. Ciobanu used generative AI to help virtually try on clothes, while Tao developed a ROS-based robot to map the inside of a home to help find things.

Harnessing LLMs for Robots Niewinski builds custom applications for robotics at his Armoury Labs business in Waterloo, Ontario, Canada, where he uses the NVIDIA Jetson platform for edge AI and robotics, creating open-source tutorials and YouTube videos following his experiences.

He built his interactive GLaDOS robot to create a personal assistant for himself in the lab. It handles queries using Transformer-based speech recognition, text-to-speech, and large language models (LLMs) running onboard an NVIDIA Jetson AGX Orin, which interfaces with a robot arm and camera for interactions.

GLaDOS can track his whereabouts in the lab, move in different directions to face him and respond quickly to queries.

I like doing things with robots that people will look at and say it's not what they had immediately expected, he said.

He wanted the assistant to sound like the original GLaDOS from Portal and respond quickly. Fortunately, the gaming company Valve has put all of the voice lines from Portal and Portal 2 on its website, allowing Niewinski to download the audio to help train a model.

Using Jetson, your average question-and-answer stuff runs pretty quick for speech, he said.

Niewinski used NVIDIA's open-source NeMo toolkit to fine-tune a voice for GLaDOS, training a spectrogram generator network called FastPitch and HiFiGAN vocoder network to refine the audio quality.

Both networks are deployed on Orin with NVIDIA Riva to enable speech recognition and synthesis that's been optimized to run at many times the real-time rate of speech, so that it can run alongside the LLM while maintaining a smooth, interactive delivery.

For generating realistic responses from GLaDOS, Niewinski uses a locally hosted LLM called OpenChat that he runs in Docker from jetson-containers, saying that it was a drop-in replacement for OpenAI's API. All of this AI is running on the Jetson module, using the latest open-source ML software stack built with CUDA and JetPack.

To enable GLaDOS to move, Niewinski developed the interactions for a Unitree Z1 robotic arm. It has a stereo camera and models for seeing and tracking a human speaking and a 3D-printed GLaDOS head and body shell around the arm.

Trying on Generative AI for Fashion Fit Winner Ciobanu, based in Romania, aimed to improve the virtual clothing try-on experience with the help of generative AI, taking a top prize for his EdgeStyle: Fashion Preview at the Edge.

He used AI models such as YOLOv5, SAM and OpenPose to extract and refine data from images and videos. Then he used Stable Diffusion to generate the images, which he said was key to achieving accurate virtual try-ons.

This system taught the model how clothes fit different poses on people, which he said enhanced the realism of the try-ons.

It's quite handy as it allows users to see how clothes would look on them without actually trying them on, said Ciobanu.

The NVIDIA JetPack SDK provided all the tools needed to run AI models smoothly on the Jetson Orin, he said.

It's super-helpful to have a stable set of tools, especially when you're dealing with AI tech that keeps changing, said Ciobanu. It really cut down on the time and hassle for us developers, letting us focus more on the cool stuff we're building instead of getting stuck on tech issues.

Finding Lost Items With Robot Assistance Winner Tao, based in Ontario, Canada, created a robot to lessen the burden of searching for things lost around the house. His An Eye for an Item project took top honors at the Hackster challenge.

Finding lost objects is a chore, and recent developments in zero-shot object detection and LLMs make it feasible for a computer to detect arbitrary objects for us based on textual or pictorial descriptions, presenting an opportunity for automation, said Tao.

Tao said he needed robot computing capabilities to catalog objects in any unstructured environment - whether a living room or large warehouse. And he needed it to also perform real-time calculations for localization to help with navigation, as well as running inference on larger object detection models.

Jetson Orin was a perfect fit, supporting all functionality from text and image queries into NanoDB, to real-time odometry feedback, including leveraging Isaac ROS' hardware-accelerated AprilTag detections for drift correction, he said.

Other winners of the AI Innovation Challenge include:

George Profenza, Escalator people tracker, 2nd place, Generative AI Applications category

Dimiter Kendri, Cooking meals with a local AI assistant using Jetson AGX Orin, 3rd place, Generative AI Applications category

Vy Phan, ClearWaters Underwater Image Enhancement with Generative AI, 2nd place, Generative AI Models category

Huy Mai, Realtime Language Segment Anything on Jetson Orin, 2nd place, Generative AI Models category

Fakhrur Razi, Autonomous Int

LINK:	https://blogs.nvidia.com/blog/glados-robot-hackster/...
	See more stories from nvidia

More from Nvidia