Embodied Representation: Robotics Without a Ruler

Conventional robotic navigation relies heavily on “sense-plan-act” loops tethered to external units like meters. If a sensor is bumped or a motor’s torque changes, the strict metric calibration fails, often resulting in crashes. In contrast, biological agents operate perfectly without knowing what a meter is, evaluating distances implicitly based on what their muscles need to do to reach a target. In this session, Levi Burner introduces Embodied Visuomotor Representation, detailing how robots can mathematically infer distance in action-implied units, successfully navigating complex tasks without any prior hardware calibration.

Key Takeaways

• Calibration breeds fragility: Depending on external units demands sub-millimeter precision from expensive stereo cameras or LiDAR. If these degrade or assumptions shift (e.g., motor wear), the underlying mathematical models collapse.
• Time-to-contact acts as a visual anchor: Inspired by the biological “tau theory,” a robot can measure the expansion rate of an approaching object to extract an exact time-to-contact state, sidestepping the need for absolute depth measurements.
• Motor outputs define the scale: By double-integrating the raw forces or torques sent to its own motors, a robot generates a self-consistent physical model. The exact scaling factor converting these embodied signals to meters remains permanently unknown—and irrelevant.
• Unknown variables cancel out: When action-driven models are fed back into the closed-loop control system, the unknown calibration multipliers naturally cancel out in the mathematics, guaranteeing stable target convergence.
• Uncalibrated oscillation solves sizing: By quickly oscillating in front of an obstacle, a robot can actively measure environmental features strictly as a ratio of its own body width, allowing it to accurately assess whether it fits through a doorway without knowing how wide it is.

About the Research

This approach directly challenges the passive “camera-as-a-passenger” paradigm established in early computer vision literature. Drawing on foundational computational neuroscience, this work implements internal feedback pathways that mimic the predictive gaze of dragonflies and the gap-sizing oscillation behaviors observed in gerbils.

As detailed in the session, practical tests proved the theory’s robust adaptability. When the control gain (the strength of the robot’s motors) was secretly doubled mid-experiment, traditional metric-based controllers completely destabilized and crashed. However, the embodied representation algorithm automatically absorbed the physical change as a simple adjustment to its internal unit scale, seamlessly settling at the target.

What This Means for Robotics

Shifting away from explicit metric modeling allows hardware to become drastically more resilient in chaotic, real-world deployments. This is particularly relevant for planetary exploration or automated driving assistance systems, where real-time recalibration is impossible if a sensor shifts or sustains damage.

By building computational representations based solely on internal body signals and active vision feedback, robots can learn to interact safely with their environments within seconds of activation. This opens the door to creating generalized control software that can be blindly deployed onto diverse robot frames without requiring manual system identification or physics engine tuning.

Resources

• Speaker Paper: Embodied Visuomotor Representation