Section 45.1: Wheeled, legged, and hybrid robots

"Mobility is morphology plus a contract with the ground."

A Builder's Locomotion Notebook

A quadruped that flawlessly clears a flight of stairs in the demo reel drains its battery pack more than twice as fast as a cheap wheeled cart on a flat hospital corridor. That single mismatch, not the stair-climbing feat, decides whether the deployment survives contact with a real building. As Figure 45.1A illustrates, different mobile bodies buy different feasible contact patterns, sensing needs, and recovery margins, which is exactly what the scoring framework below makes quantitative. For a mission class \(m\), terrain distribution \(\mathcal{T}\), and morphology candidate \(r\), a useful design score is \(J(r) = w_t T(r, \mathcal{T}) + w_e E(r, \mathcal{T}) + w_s S(r, \mathcal{T}) + w_r R(r, \mathcal{T})\), where the terms represent traversal time, energy, slip or fall risk, and recovery burden. The right body is the one that minimizes the mission score, not the one with the most dramatic demo.

Wheels obey nonholonomic constraints, such as \(\dot y \cos \theta - \dot x \sin \theta = 0\), which make them efficient on prepared surfaces but weak on discontinuous footholds. Legged bodies trade efficiency for contact flexibility. Hybrids buy mode switching, but they also add controller complexity, extra failure modes, and more sim-to-real transfer surface (the total set of modeled behaviors that must match physical reality before a policy trained in simulation will work on hardware). The counterintuitive result: the "smarter" legged policy consumes more energy on flat hospital corridors than the "dumber" wheeled one. Every step cycle lifts and replaces mass against gravity, whereas rolling converts forward momentum almost for free. On a 200-meter flat run, a quadruped expends roughly 2.2 times the energy of a comparably sized differential-drive robot (a wheeled base that steers by varying the relative speed of its left and right wheels rather than using a separate steering mechanism) covering the same distance.

The nonholonomic constraint, from math to motion

Why this matters in embodied AI. The nonholonomic constraint is not merely a math convenience; it represents a physical commitment. A wheeled robot cannot instantaneously move sideways because tires resist lateral slip through friction. In real deployments this matters acutely: a robot that cannot sidestep cannot correct a lateral position error without a wider turning arc, so narrow corridors and cluttered environments shrink the recoverable region. Planning algorithms that ignore this constraint produce trajectories the physical robot cannot follow, causing drift and localization failures.

How the constraint operates. Rolling without slip forces each wheel's contact velocity to zero perpendicular to its axis, pinning the instantaneous center of rotation to a line through the rear axle. Reachable velocity thus spans only two dimensions inside a three-dimensional pose space. A planner reaches lateral targets by sequencing arcs and in-place rotations, and the controller must track that curvature-constrained path within the surface's traction limits.

Theory

Figure 45.1.1 frames the whole process as an inspectable loop: observe the terrain, model contact affordance and mission cost, act by picking a body or support mode, then verify against completion, slip, energy, and recovery before feeding the result back into the next observation. A practical mobility stack separates three models. The geometric model asks whether the body can physically fit and place support contacts. The dynamic model asks whether required forces and torques are feasible. The autonomy model asks whether perception, state estimation, and control can keep the body inside that feasible set when the world deviates from the nominal plan.

For wheeled robots, the central quantity is curvature and traction margin. For legged robots, it is the reachable contact set and recoverable center-of-mass motion. For hybrids, it is the switching policy between support modes and the price of switching too late.

A graduate-level design habit is to compare morphologies on one scenario panel: ramps, stairs, gaps, debris, low-friction patches, narrow aisles, and payload variations. A single average speed number hides the actual body-environment contract. To see why that scenario panel has to be concrete rather than aspirational, it helps to build the smallest artifact that forces every morphology onto the same numbers.

Worked Example

The first useful artifact is not a render. It is a morphology scorecard built on the exact terrain panel that the deployment team expects to face.

from math import inf

candidates = {
    "wheeled": {"time_s": 62, "energy_kj": 1.8, "slip_events": 5, "recoveries": 7},
    "legged": {"time_s": 81, "energy_kj": 3.9, "slip_events": 1, "recoveries": 2},
    "hybrid": {"time_s": 70, "energy_kj": 3.1, "slip_events": 2, "recoveries": 3},
}

weights = {"time_s": 0.04, "energy_kj": 0.25, "slip_events": 2.0, "recoveries": 1.5}

scores = {}
for name, vals in candidates.items():
    scores[name] = round(sum(weights[k] * vals[k] for k in weights), 2)

winner = min(scores, key=scores.get)
print(scores)
print({"selected": winner, "score": scores[winner]})

Expected output interpretation. The legged body wins here because the scenario panel prices recovery burden and slip heavily: the wheeled robot's 7 recovery events alone contribute 10.5 points to its score, more than the legged candidate's entire total (9.21). A pure wheeled platform is faster and cheaper in energy, but it pays too much in recoveries and slip on this discontinuous-terrain panel. In practice, this ranking is a property of the chosen weights, not a universal verdict: a panel that priced runtime more heavily, or one drawn from a smoother terrain distribution, would typically favor the wheeled or hybrid candidate instead. The hybrid comes second here specifically because it inherits some of both bodies' costs (moderate slip, moderate recovery burden) without winning outright on either axis.

Practical Recipe

Once the scorecard and its supporting libraries can rank bodies on paper, the remaining risk is that the paper terrain never matched the field, so the recipe below turns the scoring habit into a sequence of measurements taken against the real deployment.

1. Measure the real terrain distribution before picking a body.
2. Record which tasks require continuous rolling, stepping, kneeling, bracing, or contact-rich manipulation while moving.
3. Compute mission cost with the same metric code for every morphology candidate.
4. Stress the winner with the most probable field perturbations: wheel slip, missing foothold, load shift, and degraded localization.
5. Save one decision card with geometry, dynamics, telemetry assumptions, and failure thresholds.

A robot that performs flawlessly in simulation but stalls on its first real kerb has not been validated: it has been protected from the truth. Every body choice exposes its hidden middleware and estimator implications. Wheels favor odometry and traction models. Legs require contact estimation, gait timing, and state estimation through intermittent support. Hybrids require a mode estimator that can explain why the robot changed support strategy.

This is also the right place to connect morphology to downstream chapters. The same body choice that helps locomotion may hurt manipulation reach, teleoperation ergonomics, or whole-body safety. That coupling is why embodied systems resist clean organizational boundaries.

When a morphology argument fails, isolate whether the error came from environment assumptions, force or torque feasibility, localization fragility, or operator intervention burden. Different bodies fail for different reasons, and the mitigation route changes with the reason.