By Mark Garo Ansell, Co-Founder, Sidekick Robotics, Inc.
While the broader market discusses the potential of Physical AI, the team at Sidekick Robotics has been solving the actual plumbing required to make robots move with human-like fluidity. The challenge was not just building the Sidekick AI brain, it was solving a high-speed negotiation inside the computer that allows the robot to think and act at the same time.
01The Problem: The Compute War
In a modern robot, two fundamentally different systems compete for the same GPU resources simultaneously:
Brain (AI)
The AI model that perceives the environment, reasons about the task, and decides what the robot should do next. It requires sustained GPU compute to run and cannot be interrupted mid-cycle without corrupting its output.
Sustained compute bursts
Learning (Reinforcement Learning)
The system that observes outcomes, evaluates whether the robot succeeded, and uses that signal to improve future behavior. It must be responsive enough to capture positive outcomes the instant they happen.
Time-sensitive, outcome-driven
The conflict: When the Brain is running, it occupies the compute resources the Learning system needs to observe and record outcomes. The Learning system is forced to wait, and any positive outcome that occurs during that window is lost. You cannot learn from a moment you missed.
02Sidekick's Solution: The GPU Handshake
We developed a bespoke “handshake” protocol that allows the Brain and the Learning system to share GPU resources without blocking each other. Instead of routing communication through the conventional stack, the two systems now exchange information through the Sidekick Handshake Layer, a proprietary intermediary purpose-built for this problem.
Architecture
Brain (AI)
VLA Foundation Model
Sidekick Handshake Layer
Proprietary GPU protocol
Learning (RL)
Reinforcement Learning Agent
The Brain and the Learning system communicate through the Sidekick Handshake Layer rather than through the conventional stack. The result is that neither system has to wait for the other. No missed moments.
Think of it like the difference between watching a game live and watching a replay. By the time the replay reaches you, the moment has passed. The Sidekick Handshake Layer lets the Learning system see what the Brain is doing live, as it happens.
This removes layers of scheduling latency that would otherwise be invisible in conventional software benchmarks but catastrophic for a learning system that must catch successes in real time.
03The Result: 70x Faster Performance
70x
Processing Speed
This is not a minor optimization. This is a structural change in how the robot's compute pipeline operates, and we saw its power in a concrete way during recent training runs.
“At a critical moment in a task, the robot's brain had a breakthrough. Because the handshake was so fast, the robot was able to catch that fleeting success and learn from it instantly. Without this speed, the robot would have physically moved past the goal before its brain even realized it had succeeded.”
Mark Garo Ansell
Co-Founder, Sidekick Robotics
The ability to capture and reinforce these fleeting positive outcomes is foundational to how reinforcement learning from human feedback works in practice. Speed is not just a performance metric, it is the mechanism by which learning becomes possible at all.
Further Reading
Learn more about how Sidekick Robotics is implementing Reinforcement Learning from Human Feedback, the feedback loop that transformed language models, now for physical robots.
04Why It Matters
For Sidekick Robotics, this is the bridge between a laboratory experiment and a reliable partner.
By optimizing the compute traffic jam, we have given Sidekick Robots the ability to think, act, and learn at the same time on the same high performance standard compute.
This enables scalability of our system without custom silicon that was thought to be required by some in the industry. We are excited about what this enables for Sidekick Robotics and the 70x improvement enabled by the Sidekick GPU Handshake.
Media Contact: founders@sidekickrobotics.ai