01 / OVERVIEW
Intelligence needs a safety envelope.
A reasoning system connected to a body inherits an entirely different set of consequences.
Research into Unitree Go2 and G1 robotic platforms examines the intersection of network security, embedded systems, AI-driven planning, and physical actuation. The central question is whether safety survives when one of those layers behaves unexpectedly or is compromised.
The source research describes coordinated disclosure. This overview focuses on the systems lessons.
02 / SYSTEM BOUNDARIES
Follow the path to action.
A robot is a connected system. Its app, radio, cloud services, internal buses, sensors, policy layers, and actuators share responsibility for what happens in the physical world.
When each layer assumes another layer has already enforced safety, no layer may own the invariant that actually matters.
03 / THE FAILURE PATTERN
Three conditions.
One physical consequence.
Untrusted perception
External prompts, networks, sensors, or peers can influence a plan.
Physical action authority
A plan can reach motion or force-producing functions.
Mutable safety controls
Safety checks share a trust boundary with the system they constrain.
These conditions motivate evaluating embodied AI as a complete system. Model refusal alone cannot establish that the surrounding hardware and software are safe.
04 / DESIGN IMPLICATIONS
Move safety below language.
The model may propose. A trusted controller must enforce the bounds.
The research argues for typed plans, explicit policy gates, independently enforced limits on motion and force, fail-closed behavior, and locally interruptible action. Platform security and physical safety need to reinforce one another.
Our embodied reasoning work investigates how to test those boundaries and develop evidence for the controls that depend on them.
Collaborate on embodied systems research