|
Listen to this article  |
VidGen-1 applies generative AI to generate video sequences for training autonomous systems. Source: Helm.ai
Training machine learning models for self-driving cars and mobile robots is often labor-intensive, requiring humans to annotate huge volumes of images and then supervise and verify the resulting behavior. Helm.ai says its approach to artificial intelligence is different. The Redwood City, California-based company last month unveiled VidGen-1, a generative AI model that generates realistic video sequences of driving scenes.
“By combining the deep teaching technology we have developed over the years with our in-house innovation in generative DNNs, [deep neural network] “This architecture enables a highly effective and scalable way to produce realistic AI-generated video,” said Vladislav Voroninski, co-founder and CEO of Helm.ai.
“Generative AI helps with scalability and tasks where there isn't one objective answer,” he said. Robot Report“It's non-deterministic, it looks at a distribution of possibilities, and it's important for resolving edge cases where traditional supervised learning approaches don't work. The ability to annotate data doesn't help with VidGen-1.”
Register now.
Helm.ai bets on unsupervised learning
Founded in 2016, Helm.ai develops AI for advanced driver assistance systems (ADAS), Level 4 self-driving cars, and autonomous mobile robots (AMRs). The company previously announced GenSim-1 for AI-generated and labeled images of vehicles, pedestrians, and road environments that can be used for both predictive tasks and simulation.
“We bet on unsupervised learning with a foundational model of segmentation, which is a world first,” Voroninsky says. “We're building high-end driver assistance models now, but that framework should work regardless of whether the product requires Level 2 or Level 4 autonomous driving. The workflow is the same.”
Helm.ai says VidGen-1 enables it to cost-effectively train models with thousands of hours of driving footage, enabling simulations that mimic human driving behavior across different scenarios, geographies, weather conditions and complex traffic dynamics, the company said.
“This is a more efficient way to train large models,” Voroninski said. “VidGen-1 can produce highly realistic videos without incurring prohibitive costs in computing.”
How can generative AI models be evaluated? “We have fidelity metrics that tell us how well the model approximates a target distribution,” Voroninski answered. “We have a large collection of real-world video and data, and we have models that generate data from the same distribution for validation.”
He compared VidGen-1 to large-scale language models (LLMs).
“Predicting the next frame of a video is similar to predicting the next word in a sentence, but at a much higher dimensional level,” Voroninski added. “Generating realistic video sequences of driving scenes is the most advanced form of prediction for autonomous driving, because it accurately models what the real world looks like and includes both intent prediction and path planning as implicit subtasks at the top level of the stack. This capability is critical for autonomous driving, because driving is fundamentally about predicting what will happen next.”
VidGen-1 can be applied to other domains
“Tesla may be doing a lot of work in-house on the AI side, but many other auto OEMs are ramping up their efforts,” Voroninsky said. “VidGen-1's customers are these OEMs, and this technology could help them better compete with the software they develop for consumer cars, trucks and other self-driving vehicles.”
Helm.ai says its generative AI technology offers high accuracy and scalability with a low compute profile. Helm.ai claims that VidGen-1 will help bridge the gap between simulation and reality, or “sim2real,” as it supports the rapid generation of assets in simulation with realistic behavior.
Voroninsky added that Helm.ai's models can be applied not only to generating videos for simulations, but also at lower levels of the technology stack, which he said could be used in AMRs, autonomous mining vehicles, drones and more.
“Generative AI and generative simulation will be a huge market,” Voroninski said. “Helm.ai is well positioned to help automakers reduce development time and costs while still meeting production requirements.”
