Part 5: Robotics in the Real World

Welcome to Part 5 of "Robotics Zero to Hero." We've covered the clean mathematics of kinematics and dynamics. But moving from a perfect Python simulation to physical hardware introduces an entirely new adversary: reality.

The Reality Gap

In simulation, joints go exactly where commanded. In reality, several effects open up the Sim-to-Real Gap — formally, a distribution shift between the simulator's dynamics $p_{sim}(\mathbf x'\mid\mathbf x,\mathbf u)$ and the world's $p_{real}(\mathbf x'\mid\mathbf x,\mathbf u)$:

1. Friction and stiction. A realistic joint torque includes Coulomb, viscous, and Stribeck terms:

$$\tau_f(\dot q) = \underbrace{f_c\,\mathrm{sgn}(\dot q)}_{\text{Coulomb}} + \underbrace{f_v\,\dot q}_{\text{viscous}} + \underbrace{(f_s - f_c)\,e^{-(\dot q/v_s)^2}\,\mathrm{sgn}(\dot q)}_{\text{Stribeck (stiction)}}$$

The $\mathrm{sgn}(\dot q)$ discontinuity at $\dot q=0$ is what makes friction so hard to model and control near rest. 2. Backlash. Gear "play" means the motor turns a small dead-band angle before the link moves — a hysteresis nonlinearity no rigid model predicts. 3. Sensor noise. Encoders and cameras feed the controller noisy, quantized, latent data. 4. Payload variation. Pick up a heavy object and the inertia matrix $M(\mathbf q)$ from Part 4 changes instantly — your carefully-identified model is now wrong.

Closing the gap: competing strategies

There is no single fix; the field uses a portfolio of methods, each with a characteristic weakness:

Domain randomization deserves special mention: by exploiting the linearity-in-parameters structure from Part 4, we can either identify $\boldsymbol\Theta$ or randomize over it. The seminal result (Tobin et al., arXiv:1703.06907) showed a vision model trained purely on randomized simulation transferring to the real world with no real images at all.

Internet AI vs. Embodied AI

As Professor Jitendra Malik points out, much of modern AI is "Internet AI" — it learns passively from massive static datasets of text and images scraped from the web (think large language and image models).

Robotics demands "Embodied AI." An embodied agent learns by actively perceiving, moving, and interacting in a 3D world, where its own actions change the data it sees — a feedback loop Internet AI never experiences. Because we cannot train millions of times on physical hardware without destroying the robot, bridging the gap relies on high-fidelity simulators (e.g. AI Habitat, Isaac Sim, MuJoCo) where agents practice in photorealistic, physically-plausible worlds before deployment.

Industrial Cobots and Safety Protocols

Historically, industrial robots were massive, fast, and dangerous — caged to keep humans out. The modern paradigm is the Collaborative Robot (Cobot), designed to share space with people. This requires tight hardware/software integration with safety standards (ISO/TS 15066):

• Force/torque limiting. Cobots watch joint torques; an unexpected spike (a collision with a human) triggers an immediate halt. This is enabled by accurate dynamics: to detect an external torque you must predict the expected internal torque $\hat{\boldsymbol\tau} = M\ddot{\mathbf q} + C\dot{\mathbf q} + \mathbf g$ and flag the residual.
• Speed and separation monitoring. Using LiDAR or depth cameras, the robot slows as a human approaches and stops within a defined radius.

Focus on the Octopus: Mechanical Design and Aging

Building a metallic continuum octopus forces us to rethink mechanical design.

A standard robot puts a heavy motor at every joint. Put a motor in every segment of a tentacle and it becomes too heavy to lift itself — the payload-to-weight ratio collapses.

The cable-driven approach

We instead use a cable-driven (tendon-driven) architecture: heavy motors stay in the stationary base (the head), and cables run through the hollow segments to the tips. Pulling specific cables induces curvature, bending and twisting the tentacle while keeping it light and fast.

The price is a brutal amplification of the sim-to-real gap. The map from actuator space (cable lengths $\boldsymbol\ell$) to configuration space (curvatures $\mathbf q$) is nonlinear and state-dependent:

$$\boldsymbol\ell = h(\mathbf q) + \underbrace{\boldsymbol\delta_{stretch}(\boldsymbol\tau)}_{\text{cable elongation}} + \underbrace{\boldsymbol\delta_{friction}(\mathbf q,\dot{\mathbf q})}_{\text{capstan friction}}$$

Cables stretch under load and rub against segment walls (capstan-effect friction grows exponentially with the wrap angle), so the same cable pull produces different curvature depending on pose and history.

Aging and recalibration

Hardware degrades. Steel tendons slowly stretch after thousands of actuation cycles, and friction coefficients drift as internal surfaces wear — hardware aging. A controller that assumes a brand-new robot will see its accuracy decay over months.

The fix is continuous recalibration: comparing the model-predicted motion against the measured motion (vision or proprioception) and updating the internal parameters $\boldsymbol\delta_{stretch}, \boldsymbol\delta_{friction}$ online — a recursive least-squares / Kalman-style estimator running in the background. This is adaptive control (above) applied to an aging body, and it is what keeps millimeter precision over the robot's lifetime.

High-dim vs. low-dim, restated for calibration

Calibration scales badly with dimension. A 6-DOF arm has a handful of kinematic parameters to identify; a tentacle modeled as $N$ segments has $O(N)$ curvature maps, each with stretch and friction terms — and they interact. The more degrees of freedom, the more parameters drift, and the more essential continuous self-calibration becomes. High-dimensional robots are not just harder to control — they are harder to keep honest.

In Part 6 we introduce Singular Perturbation, the mathematical tool for managing the multi-time-scale dynamics this hardware produces.

Further reading: Tobin et al. (arXiv:1703.06907) on domain randomization; Siciliano & Khatib (eds.), "Springer Handbook of Robotics" for friction/identification; the capstan equation for tendon friction.