Stanford Robotics Seminar ENGR319 | Spring 2026 | Mechanical Intelligence in Locomotion

Key Concepts

• Meso-scale Robots: Robots weighing approximately 1 kg, filling the research gap between micro-robots (<1g) and large-scale robots (>10kg).
• Morphological Intelligence: The ability of a robot to navigate complex environments through passive physical responses governed by its body structure and physical constraints, rather than active computational feedback.
• Noise-Dominated Regime: An environment where terrain interaction forces are unpredictable due to the robot's size relative to environmental obstacles (interacting with ~10 objects simultaneously).
• Coasting Number: A dimensionless number derived from fluid mechanics used to quantify the role of inertia versus friction/viscosity in locomotion.
• Embodied Intelligence: A framework defined as the sum of Morphological Intelligence and Computational Intelligence.
• Shannon’s Information Theory Analogy: Using the principles of signal transmission (redundancy and coding) to explain how redundant physical structures (e.g., many legs) allow for reliable locomotion without feedback.

1. The Case for Meso-scale Robots

Meso-scale robots are essential for applications where large robots are too clunky and small robots lack the power to overcome obstacles.

• Applications:
  • CBRN Threat Detection: Chemical, biological, nuclear, and radioactive detection in confined spaces (Market projection: $21.5B by 2028).
  • Precision Agriculture: Weed removal and disease detection without damaging crops (Market projection: $21.9B by 2031).
  • Indoor/Safety: Smart furniture, mining safety, and forestry management.

2. The Challenge of the "Noise-Dominated Regime"

At the meso-scale, robots interact with approximately 10 objects at once, each with a weight comparable to the robot. This creates a "noise-dominated" environment where terrain reaction forces are nearly impossible to predict using standard inertia-driven models.

3. Morphological Intelligence: A Theoretical Framework

The speaker proposes that computational intelligence (sensing and responding) is not the only way to achieve reliable locomotion.

• The Analogy: Just as Shannon proved that reliable signal transmission over noisy channels is possible through redundancy rather than just feedback, a robot with sufficient morphological redundancy (e.g., 16 legs vs. 6) can achieve reliable, predictable locomotion over complex terrain without feedback.
• Experimental Evidence: Experiments with 6- to 16-legged robots showed that as the number of legs increases, the variance in arrival time decreases, converging to a "guaranteed" locomotion speed regardless of terrain heterogeneity.

4. Beyond Bio-Inspiration: Asymmetry and Optimization

The speaker argues that "taking inspiration from biology" should not mean mere imitation, but rather knowing when to transcend biological constraints.

• Asymmetry: While biological systems often rely on bilateral symmetry, graph optimization reveals that asymmetry can lead to faster locomotion.
• Case Study: A hexapod robot with an asymmetric gait (turning clockwise for 3/4 of a cycle and counter-clockwise for 1/4) achieved 50% higher speeds than symmetric counterparts. Some motors were even replaced with rigid parts, proving that non-actuated components can contribute to efficient movement.

5. Methodologies for Adaptive Control

• Skink Robot (Sensing Modality): By studying the torque distribution on the body of a skink-inspired robot, the team developed a method to classify terrain depth (granular media) with 90% accuracy.
  • Key Insight: As limb size decreases, the optimal sensing modality shifts from the legs to the body.
• Graph Optimization: By treating contact states as vertices and displacement as edges, the team converted the complex problem of locomotion into a graph optimization problem, allowing for the discovery of non-intuitive, high-performance gaits.

6. Synthesis and Conclusion

The speaker concludes that the future of robotics lies in the co-design of morphological and computational intelligence.

• Functional Equivalence: There is a trade-off between physical structure and computational effort; a 16-legged robot with no feedback can perform as well as a 12-legged robot with low-frequency feedback.
• Actionable Insight: To build effective meso-scale robots, engineers should quantify the "cost of transport" for both mechanical and computational components. By balancing these, one can generate "embodied intelligence on demand," tailored to specific speed and robustness requirements.

"It is the asymmetry in the morphological intelligence plus the asymmetry in the computational intelligence that can give us the reliable locomotion." — Speaker