Our team had the opportunity to attend AI for Good's session "Robotic Foundation Models" which validated the core thesis of our company, building on top of recent progress in the research community that has made it possible to train large neural network models for robotic control that enable tasks that were previously impossible, by transferring semantic knowledge from web-scale pretraining and combining it with physical understanding learned from robot data.
Image Credit: Sergey Levine, Physical Intelligence
By integrating tightly with robotic foundation model progress, we continue to build the best possible solutions for customers seeking physical labor automation.
From LLMs to VLMs to VLAs
VLAs (Vision-Language-Action models) are advanced LLMs that can also control robotic actions, representing the cutting edge of robotic foundation models.
Image Credit: Sergey Levine, Physical Intelligence
Key Innovation: 2nd generation VLAs add a dedicated continuous output mechanism that can specialize for motor control, like a "virtual motor cortex" — enabling robots to perform complex physical tasks with unprecedented precision.
Validation from a Leading Expert
Sergey Levine is Associate Professor at UC Berkeley and head of the Robotic Artificial Intelligence and Learning Lab and co-founder of leading Robotic Foundation Model company Physical Intelligence.
With decades of experience, he is a leader in the field of algorithms that can enable autonomous agents to acquire complex behaviors through learning.
3 Key Quotes that Validate Our Work at Sidekick Robotics
"Even with the pi0.5 model, the absolute performance is pretty good but not at the level that we want from a practically deployed model."
— Professor Sergey Levine, UC Berkeley
"A lot more research needs to be done to get them to the 99.9% reliability to get them to solve long horizon tasks."
— Professor Sergey Levine, UC Berkeley
"Training VLAs with reinforcement learning remains the cutting edge."
— Professor Sergey Levine, UC Berkeley
The Algorithm Layer Advantage
Foundation Model Layer
Current robotic foundation models achieve impressive capabilities but fall short of production-ready reliability.
Algorithm Layer (Sidekick)
By focusing on the algorithm layer with reinforcement learning, Sidekick Robotics delivers the reliability needed for commercial deployment.
This is exactly what Sidekick Robotics focuses on: the algorithm layer that bridges the gap between impressive research demonstrations and commercially viable robotic systems.
Looking Ahead
We're excited to see the research community continue innovating when it comes to developing memory, long-horizon reasoning, and more. It's such an exciting time and great validation about the work we're doing applying reinforcement learning algorithms in post-training.
The convergence of robotic foundation models and advanced algorithm layers represents the future of physical AI — and Sidekick Robotics is at the forefront of making this future a reality.
About Sidekick Robotics
Sidekick Robotics is creating the AI Sidekick for Physical Work -- robots that learn and perform the complex tasks that keep hospitals and care communities running. We're tackling the healthcare labor shortage by reducing burnout, turnover, and rising costs, so caregivers can focus on what matters most: patients.
Media Contact: founders@sidekickrobotics.ai