Measuring Slip the Same Way in Simulation and on Hardware for Wheeled-Legged Robots

Wheeled-legged robots, which are quadrupeds with motorized wheels instead of point feet, can roll across flat ground at a fraction of the energy cost of walking and still step over curbs. Rolling only pays off while the wheels keep traction. Simulators are poor at modeling exactly the thing these robots depend on: the small, stick-slip skidding of a wheel on wood, mud or gravel. Animation researchers face a related problem called "foot sliding," where generated characters' feet glide over the floor. Their standard fix is to penalize foot velocity whenever the foot touches the ground. That fix is the wrong tool for a wheel, whose whole purpose is to move while in contact.

This note proposes one quantity that applies to both cases. The slip residual is the velocity of the material point of the wheel or foot that is touching the ground, which is zero for both perfect rolling and planted stepping. It can be computed from the same signals in simulation and on hardware. We use it in two ways. First, it serves as the safety axis of a single locomotion policy that accepts a commanded slip budget and trades energy against traction on demand. Second, it serves as a diagnostic of the reality gap. Our central hypothesis is that, for wheeled-legged locomotion, the sim-to-real gap is concentrated along the slip axis. If that holds, a handful of hardware runs can recalibrate the mapping from commanded to realized slip, in place of broad and conservative randomization of simulator physics.

Core Claim

A rolling wheel's contact point is instantaneously at rest. The wheel center moves, but the rubber touching the ground does not. Zero-velocity losses from motion synthesis therefore fail on wheels for a specific reason: they are applied to the wheel frame rather than to the material contact point. The binary contact classification with velocity clamping in Combining Recurrent Neural Networks and Adversarial Training for Human Motion Synthesis and Control is one example. Applying the penalty to the correct point removes the conflict between "no sliding" and "rolling."

For contact $i$, let $v_i$ be the wheel-center velocity, $\omega_i$ the wheel angular velocity, and $r_i$ the vector from wheel center to ground contact. The vector $r_i$ uses the effective rolling radius, which changes with leg pitch and roll. The slip residual is

$$ s_i = v_i + \omega_i \times r_i , $$

and its lateral and longitudinal components are tracked separately. A stepping foot has $\omega_i = 0$, so the same expression reduces to the animation-style zero-velocity condition.

On hardware, each term is available onboard. The base state comes from an estimator. Leg kinematics give $v_i$, wheel encoders give $\omega_i$, and wheel geometry gives $r_i$. The estimator should be geometry-aware in the style of State Estimation for Hybrid Locomotion of Driving-Stepping Quadrupeds, which showed that ignoring wheel radius causes height and velocity drift. Motion capture provides ground truth for validation. Friction-Aware Safety Locomotion for Wheeled-legged Robots using Vision Language Models and Reinforcement Learning already penalized slip with a simulation-only "translation joint." Our residual is intended as the version of that quantity that can be measured on hardware.

The policy is trained as a constrained problem with the budget as an input. For each episode we sample a slip budget $b$ and solve

$$ \max_\pi \; \mathbb{E}\big[-\mathrm{CoT} - \lambda_{\mathrm{track}} e_{\mathrm{track}}\big] \quad \text{s.t.} \quad \mathbb{E}\big[\textstyle\sum_i \|s_i\|\big] \le b , $$

with $b$ observed by the policy. Here CoT is the mechanical cost of transport, $e_{\mathrm{track}}$ is velocity-command tracking error, and the constraint is enforced with a Lagrange multiplier that is itself conditioned on $b$. We choose a constraint rather than a weighted sum because sweeping a constraint level reaches non-convex parts of a trade-off curve that weighted sums miss. The file names this weighted-sum failure as an open problem in multi-objective trajectory optimization.

Research Program

The first component is the residual itself, together with a test of its hardware fidelity. On a wheeled quadruped, we compare onboard-computed $s_i$ against motion-capture-derived contact-point velocity across surfaces of differing friction. Rubber, wood, cardboard, grass and loose gravel cover the range used in the friction-aware work. If onboard $s_i$ cannot resolve slip events above estimator noise, the hardware half of the program fails. We would then know this early.

The second component is the budget-conditioned policy. It is trained in Isaac Gym with an asymmetric actor-critic: the critic sees privileged friction and contact forces, and the deployed actor uses proprioception plus $b$. This setup follows the recipe in Reinforcement Learning for Blind Stair Climbing with Legged and Wheeled-Legged Robots and Learning Robust Autonomous Navigation and Locomotion for Wheeled-Legged Robots. Neither of those works exposes a traction/efficiency trade-off as an input. The stair-climbing work also relies on a manual boolean mode switch.

We impose no mode switch. Whether the robot rolls or steps on a given patch should emerge from the budget. A tight budget on low friction should induce stepping or slow rolling. A loose budget on rubber should induce fast driving. This parallels how gait emerges from kinematic utility in Whole-Body MPC and Online Gait Sequence Generation for Wheeled-Legged Robots, with traction rather than reach as the driving signal.

The third component is gap localization and calibration. We trace the trade-off front (CoT, tracking error, realized slip) as $b$ varies, first in simulation and then on hardware. Our hypothesis predicts that the hardware front is displaced mainly along the slip axis, with CoT and tracking roughly preserved at matched behavior. The reasoning is that rigid-body engines treat wheels as idealized rolling constraints, whereas energy and tracking depend mostly on mass and actuator properties, which are easier to identify.

If the hypothesis holds, calibration needs only a one-dimensional monotone map per surface class, from commanded $b$ to realized hardware slip. We fit it from roughly ten short runs and invert it at deployment so that the operator's requested slip is the slip actually obtained. This step is deliberately much smaller than learning a full action transformer, as in Reinforced Grounded Action Transformation for Sim-to-Real Transfer, which grounds joint actions and was validated only between modified MuJoCo environments.

Baselines

The naive reference is a PPO policy with standard domain randomization over friction, mass and latency and no slip term. Its reward is command tracking plus an energy penalty, trained on the same terrains, with the same compute and the same number of environment steps. It tells us how much slip ordinary randomization already suppresses, and at what energy cost.

The animation-transfer baseline is identical except for one term: a zero-velocity penalty on the wheel frame during contact, the direct port of motion-synthesis sliding losses. We expect it to suppress rolling and raise CoT toward walking levels. If it does not, our premise about why such losses fail on wheels is wrong.

The strongest learned competitor is a friction-conditioned policy in the style of the friction-aware work. It receives a friction estimate and is penalized for slip with the simulation-only translation-joint measure. We give it ground-truth friction plus noise matched to the reported vision estimator error, an RMSE of about 0.25. On its two-wheeled pendulum it reports 17/17 successful hardware tracking trials across five surfaces, while LQR failed completely on wood. We do not compare those numbers directly because the platform differs. This baseline conditions on the environment, whereas ours conditions on the desired trade-off.

The strongest multi-objective competitor is a sweep of separately trained policies, one per scalar slip-penalty weight. We use five to seven weights, with a total compute budget equal to our single conditioned policy and an identical tuning budget. This is the baseline we are most worried about.

For context on efficiency, we report figures from prior systems. The whole-body MPC of Whole-Body MPC and Online Gait Sequence Generation for Wheeled-Legged Robots reports CoT of about 0.1 for pure driving at 2.0 m/s and up to 85% CoT reduction relative to walking. The hierarchical system of Learning Robust Autonomous Navigation and Locomotion for Wheeled-Legged Robots reports mechanical CoT of 0.16 at 1.68 m/s. These are ANYmal-scale figures, so they serve as sanity bounds on our own platform rather than as head-to-head targets.

Beating the baselines means the following. At matched realized hardware slip, our front should reach lower CoT than the randomization baseline and the friction-conditioned baseline. Its hypervolume should be at least equal to the weight sweep's at equal compute. Across surfaces, the calibrated policy should hit the commanded slip within a tolerance we fix in advance.

Two ablations isolate the key ingredients. The first replaces the geometry-aware residual with one using a fixed nominal radius. The second removes calibration, deploying the simulator's $b$ unchanged. A third replaces constraint conditioning with weight conditioning, to test whether the non-convexity argument matters in practice.

Why not just…?

Why not just randomize friction widely and let PPO absorb slip? The file notes that broad randomization makes policies conservative. The friction-aware work also reports that randomization and adaptation baselines fell short of friction-informed control on slippery surfaces. Still, this is an empirical question for our platform. The naive baseline answers it: if it already lies on our hardware front, our contribution reduces to the measurement tool.

Why not just train a few policies with different penalty weights and pick one? This is a real threat. The weighted-sum non-convexity argument may simply not bite here if the true front is convex. We therefore treat the sweep as the primary multi-objective competitor at equal compute. If it matches our front, we will report that the gain comes from the residual and the calibration, not from conditioning. Calibration applies to the sweep too: the sweep's members can be calibrated the same way. Run-time adjustability without retraining remains a secondary advantage.

Why not just add the existing translation-joint slip penalty? It does not exist on hardware. A penalty that cannot be measured on the real robot cannot reveal or correct a sim-to-real discrepancy in slip. The friction-conditioned baseline uses this penalty, so it directly measures what hardware measurability adds.

Why not ground the simulator with a learned action transformer instead of a one-dimensional calibration? A full transformer is far more data-hungry, and its only reported validation is sim-to-sim. If our gap-localization test shows the discrepancy is not confined to slip, the one-dimensional calibration is insufficient. In that case the transformer becomes the appropriate next step, and we would say so.

Proposed Experiments

The study starts in simulation. We train in Isaac Gym on a wheeled quadruped model over flat, stepped (up to roughly 20% of leg length, matching Rolling in the Deep – Hybrid Locomotion for Wheeled-Legged Robots Using Online Trajectory Optimization), and mixed-friction terrain. Before touching hardware, we use MuJoCo with a different contact model as a sim-to-sim stand-in for "reality." Hardware experiments use a wheeled quadruped of the modified Mini Cheetah class, with motion capture over the five surfaces plus gravel. Metrics are mechanical CoT, velocity-tracking error, onboard and motion-capture slip residuals, front hypervolume, and the fraction of stance time spent rolling versus stepping.

Three checks decide the outcome. The first is whether onboard $s_i$ correlates with motion-capture slip well enough to rank slip events on hardware. The second is whether the sim-to-hardware displacement of the front is dominated by the slip axis; if CoT shifts as much as slip, the calibration story is falsified. The third is whether the calibrated budget policy achieves the requested slip on unseen surface instances while matching or beating the weight sweep at equal compute. A secondary check is that the wheel-frame zero-velocity baseline measurably reduces rolling time and raises CoT. That result would confirm, in a controlled setting, that animation-style sliding losses are mis-specified for rolling contacts rather than merely mistuned.