Robotics Motion-Control Loop with Safety Layer

A robotics motion-control loop with a safety filter, drawn as a closed loop with five blocks.

Top — Reference Policy:
- A learned policy (e.g., diffusion policy or VLA) outputs a desired joint trajectory q_des(t).

Right — Safety Filter:
- A model-based filter computes an instability score R(t) from the proposed action and current robot state.
- If R(t) exceeds a threshold, the filter falls back to a physics-guided reference (e.g., interpolation toward a safe pose); otherwise it passes the policy output through.

Bottom — Low-level Controller:
- Tracks the (possibly filtered) reference using PD + computed-torque, producing motor torques tau(t).

Left — Robot Dynamics:
- Receives tau(t), produces actual joint angles q(t) and velocities q_dot(t) via rigid-body dynamics.

Center — State Observer:
- Aggregates sensor readings into the robot state x(t) consumed by the policy and the safety filter.

A small inset graph at the bottom-right shows the instability score R(t) over time, with a horizontal threshold line.

Style: clean control-systems schematic, navy / amber palette, white background, sans-serif labels. Suitable for robotics conferences (ICRA, CoRL).