Sidekick Robotics
Safety @ Sidekick
How should a robot behave when it's still learning?
It's the question underneath Physical AI, and it's easy to get backwards. A robot that learns from its own experience will, by definition, try things no one scripted.
That's the point, and it's also the risk.
We don't ask the model to learn to be safe. We design safety as the floor the robot can't fall through.
The Architecture of Safe Autonomy
Safety is enforced beneath the intelligence, not inside it.
Learning Agent
Explores, plans, and adapts within bounded limits it cannot override.
AI Robotic Policy
Translates intent into motion using real hardware signals, not ideal assumptions.
Safety Layer
Hard limits enforced below the policy. Joint ranges, motion speeds, fault guards: immovable and always on.
The Principles We Build By
1.How the agent is constrained
Hard Limits
Beneath the policy, not inside it
Joint ranges and motion speeds are enforced below the learning agent, in a layer it cannot override. Safety isn't a behavior we hope emerges from training. It's a constraint the system is built around.
Bounded Exploration
By design
The agent is built to explore, but exploration is bounded at the source. The range of what it can attempt is capped before any command is formed, so trying something new can never mean trying something unsafe.
Trust Model
Earn trust before acting
Act only from a known-safe state. The system doesn't begin steering until it has grounding to steer from, and heavy work happens only when the robot is at rest, never mid-motion.
2.How the guards behave
Real Signals
Guards key off reality, not assumptions
Protections respond to what is actually happening on the hardware: measured motion, motor temperature, not the ideal case. Safeguards that trust the nominal case fail exactly when they matter most.
Graceful Degradation
Fail safe, not silent. And never all at once.
Heat, abnormal motion, and faults don't hit a single cliff. They escalate: warn, then pause, then stop, easing back only once conditions are verified safe. Graceful degradation beats a hard failure mid-motion, every time.
Independent Layers
No single guard is trusted alone
Limits on position, speed, temperature, and system integrity run as independent layers. If one misses, another still holds. Safety that depends on one thing working is not safety.
3.Reliability as safety
Human in the Loop
Autonomy is earned, not assumed
A person stays in the loop while the robot is learning. Autonomy is something the system grows toward through human-guided alignment, not a switch flipped on day one.
Reliability
Reliability as a safety property
A system that stalls or crashes during a motion is a safety problem, not just a bug. Stable execution, integrity checks, keeping the control loop alive: this is safety work too.
Smooth by Construction
No sudden moves
Where one motion plan hands off to the next, the system blends the seam so the robot never lurches between commands. Smoothness isn't comfort. A discontinuity is a velocity spike waiting to fault a motor.
Safety is the foundation everything else at Sidekick is built on.
Running a learning robot on real hardware is unforgiving: a single oversized command can fault a motor or damage equipment. The layered guards that prevent it, beneath the policy, on real signals, degrading gracefully, are what make it safe for a robot to explore, improve, and operate in the real world.
Learn more or get involved
Questions about our safety approach, or ready to trial our robotic intelligent solutions?