Why Bother with Adaptive Control in the Age of Robot Learning?

1. The Uncomfortable Question

If you have been following the robot learning literature over the past few years, you have seen some genuinely remarkable things: quadrupeds parkour over obstacles, humanoids walk out of labs onto outdoor terrain, dexterous hands manipulate cables and deformable objects. Almost all of these successes share a common recipe: train a policy in simulation, randomize the physics parameters over some distribution, then deploy in the real world. The results are impressive enough that one might reasonably ask:

I want to argue that the answer is yes, and that understanding why requires thinking carefully about what kind of problem the sim2real gap actually is. This is not an argument that learning is insufficient — rather, it is an argument that the two approaches are solving structurally different problems and are most powerful in combination. But to make that case, I need to first be precise about what the sim2real gap really looks like, especially for contact-rich systems.

2. The Sim2Real Gap Is Not Noise

The sim2real gap is commonly framed as a distributional shift problem: your simulator samples from one distribution of physical parameters and environments, and the real world is a sample from a different, unknown distribution. Domain randomization (DR) addresses this by broadening the training distribution so that the real world plausibly falls within its support.

This framing is useful but incomplete. The sim2real gap has structure. The errors you encounter at deployment time are not i.i.d. noise — they are systematic, correlated, and physically interpretable. Understanding this structure is what adaptive control is designed to exploit.

Systematic Inertia and Actuator Errors

Rigid body inertia tensors are notoriously difficult to measure accurately. CAD models assume uniform density distributions that do not reflect cables, wiring harnesses, PCBs, or gearbox mass. When a leg strikes the ground, the error in your inertia model produces a systematic tracking error — not random noise, but a consistent bias that depends on the current configuration and velocity.

Actuator dynamics compound this. Real motors have thermal-dependent winding resistance, back-EMF that changes with velocity, and gearbox compliance that varies with load. A motor commanded to produce 20 Nm at low speed on a cold morning may behave differently than at operating temperature. These dynamics are not typically modeled in fast-update rigid body simulators like Isaac Gym or MuJoCo, and they produce correlated errors in torque tracking across an entire joint trajectory.

Contact Model Failures

The largest source of sim2real error in legged locomotion and manipulation is contact. Standard simulators use rigid-body contact: instantaneous, perfectly stiff collisions resolved with impulse-based or penalty-based methods. Real contact is none of these things.

Real foot–ground contact involves:

• Ground compliance. Grass, gravel, and foam all deform under load. The effective contact stiffness and damping change the timing of leg loading and unloading, which directly affects gait stability.
• Foot deformation. Compliant rubber pads on robot feet deform in ways that shift the effective contact point and create a distributed pressure field, not a point contact.
• Penetration depth and hysteresis. Soft-body contacts exhibit energy dissipation that is path-dependent — loading and unloading curves differ. This is essentially invisible to a rigid-body simulator.

For dexterous manipulation, the problem is even more acute. Object contact during grasping involves multi-point contact patches, fingertip compliance, and grasp stability conditions that depend on contact normal directions that shift as the object moves.

Friction Model Failures

The Coulomb friction model — universally used in rigid-body simulation — is wrong in at least three important ways for robotics applications:

• Stribeck effect. At very low relative velocities, friction transitions from static to kinetic through a regime where friction force actually decreases with increasing velocity before increasing again. This produces stick-slip behavior that can destabilize a leg during slow stance phases.
• Friction anisotropy. Real surfaces are not isotropic. Terrain like wet grass, grooved surfaces, or sand have direction-dependent friction coefficients. A policy trained on isotropic friction will not correctly anticipate lateral slip forces on these surfaces.
• Rate-and-state friction. In granular materials and some soft surfaces, the friction coefficient depends on the recent contact history — how long the surfaces have been in contact and at what relative velocity. This is state-dependent in a way that Coulomb friction cannot capture.

3. What Domain Randomization Can and Cannot Do

Domain randomization is genuinely remarkable. If you randomize friction coefficients, ground stiffness, link masses, and motor gains over a wide enough range during training, you obtain policies that transfer to real hardware with surprising reliability. The implicit mechanism is that the policy learns to read the physical regime it is in from proprioceptive and exteroceptive observations, and adjusts its behavior accordingly. This is sometimes called implicit adaptation.

But DR has two fundamental limitations that are worth naming precisely.

Limitation 1: Coverage does not imply identification. DR ensures (roughly) that the real world falls within the support of your training distribution. But it does not help the deployed policy know where in that distribution it currently is. All environments are implicitly treated as equally likely at every time step. The policy might behave robustly on average, but it cannot optimally specialize its behavior to the current physical parameters because it does not maintain an explicit estimate of them.

This matters most when the optimal behavior differs significantly across the parameter distribution. A legged robot carrying a 5 kg payload and one carrying a 20 kg payload should use different gait strategies, different foot placement timing, different stance widths. A DR-trained policy will try to hedge between these behaviors, which is suboptimal for both.

Limitation 2: Graceful degradation is not guaranteed near distribution boundaries. Near the edges of the training distribution, the policy has seen fewer examples and its behavior becomes less reliable. More problematically, the policy has no mechanism to detect that it is near a boundary. It does not output uncertainty; it does not slow down or become more conservative; it just applies a learned function to an input that is somewhat out of distribution. This produces hard-to-predict failure modes at exactly the conditions where you most need reliable behavior.

4. What Adaptive Control Offers

Adaptive control addresses the parametric uncertainty problem directly. Rather than training to be robust against all possible uncertainties, the goal is to estimate uncertain parameters online and adjust the controller accordingly.

Consider a system with dynamics:

$$\dot{x} = f(x, u, \theta)$$

where $x$ is state, $u$ is control input, and $\theta \in \mathbb{R}^p$ is a vector of uncertain physical parameters (e.g., link masses, friction coefficients, contact stiffness). In a classical adaptive control framework, you jointly run:

1. A parameter estimator that maintains $\hat{\theta}(t)$, updated based on observed prediction errors.
2. A control law that uses $\hat{\theta}(t)$ in place of the true $\theta$, and is designed to maintain stability even when $\hat{\theta} \neq \theta$.

The stability analysis typically uses a Lyapunov function of the form:

$$V(e, \tilde{\theta}) = \frac{1}{2} e^T P e + \frac{1}{2\gamma} \tilde{\theta}^T \tilde{\theta}$$

where $e$ is the tracking error, $\tilde{\theta} = \theta - \hat{\theta}$ is the parameter estimation error, $P \succ 0$ satisfies a Lyapunov equation for the nominal closed-loop system, and $\gamma > 0$ is an adaptation gain. With the right update law for $\hat{\theta}$, one can show $\dot{V} \leq 0$, which gives stability and convergence guarantees for the joint system.

This is the Model Reference Adaptive Control (MRAC) paradigm. In practice, the guarantees come with assumptions — persistent excitation for parameter convergence, matched uncertainty structure, slow parameter variation — but the core guarantees are more tractable than those of any learned policy.

L1 Adaptive Control

A particularly attractive modern variant is L1 adaptive control [1]. Classical MRAC has a well-known tension: faster adaptation gains improve parameter estimation speed but tend to inject high-frequency content into the control signal, exciting unmodeled dynamics and potentially destabilizing the system. L1 AC resolves this tension elegantly.

The key idea is to separate estimation from control via a low-pass filter. The estimator runs at high bandwidth (updating $\hat{\sigma}(t)$, a matched estimate of unmodeled forces/disturbances, very rapidly). The control input is computed as:

$$u(s) = C(s) \, \hat{\sigma}(s)$$

where $C(s)$ is a stable low-pass filter. The filter prevents high-frequency estimation content from reaching the actuators, while still allowing the controller to reject disturbances in the frequency band where the control system has authority.

The crucial result: for a given filter $C(s)$ and adaptation rate $\Gamma$, one can derive an explicit bound on the transient deviation of the actual system from the desired reference model:

$$\|x - x_{\text{ref}}\|_{\mathcal{L}_\infty} \leq \frac{\|G(s)\|_{\mathcal{L}_1}}{\Gamma}$$

where $G(s)$ depends on the filter and system structure. This is a pre-deployment guarantee: before you turn on the robot, you know an upper bound on how far the real trajectory can deviate from the reference. For safety-critical systems, this is enormously valuable.

5. Why Contact-Rich Systems Are Special

The argument for adaptive control becomes even stronger when you focus on contact-rich systems. Three properties make these systems particularly challenging for pure learning approaches.

Contact induces switching dynamics. A legged robot alternates between contact and flight phases. Each contact event is a discrete, hybrid transition that changes the system's effective dynamics discontinuously. The stance phase has a different Jacobian, a different effective inertia, and a different set of feasible control inputs than the flight phase. Adaptive controllers designed for smooth ODEs do not directly apply here — this is an active research area involving hybrid adaptive control and reset maps.

Contact forces are unilateral and friction-constrained. Ground reaction forces at a foot must lie inside the friction cone:

$$\sqrt{f_x^2 + f_y^2} \leq \mu \, f_z, \quad f_z \geq 0$$

Whether a commanded foot force is feasible depends critically on $\mu$, the friction coefficient between foot and ground. Underestimating $\mu$ makes the controller over-conservative (wide stance, slow gaits). Overestimating $\mu$ causes slipping — a discrete, catastrophic failure. On-the-fly estimation of $\mu$ from observed foot forces and velocities is a natural adaptive estimation problem.

Manipulation contact involves unknown object properties. For dexterous manipulation of novel objects, the controller does not know:

• Object mass and inertia (needed for dynamic manipulation planning)
• Surface friction at contact points (determines grasp stability)
• Object compliance (affects deformable object manipulation)

In regrasping, in-hand manipulation, and tool use, these unknowns must be estimated online from fingertip force/torque feedback during the manipulation itself. This is an adaptive identification problem, and classical methods (recursive least squares, extended Kalman filters on physical parameters) provide both estimates and uncertainty bounds that a neural network does not.

6. Learning + Adaptive Control: Better Together

The most promising near-term direction is not "learning vs. adaptive control" but integrating both. The two approaches have complementary strengths:

• Learned policies handle high-dimensional sensory inputs, complex task structure, multi-modal behavior, and situations that are too complex to specify analytically (terrain navigation, whole-body contact planning).
• Adaptive controllers handle structured physical uncertainty at the level of forces, torques, and physical parameters — with online estimates, principled update laws, and stability guarantees.

A natural architecture: a learned policy operates at the level of task-space commands or centroidal trajectories, while an adaptive layer compensates for physical model errors at the joint/force level before those commands reach the hardware.

This is the approach I worked on in the OCRL course project at CMU: augmenting the CaJun learned centroidal locomotion controller with an L1 adaptive layer to handle unexpected payloads. The learned controller handles gait selection, balance, and high-level agility; the adaptive layer handles the structured inertia uncertainty introduced by varying payload mass. The two layers do not need to know about each other's internals — the interface is a clean torque command signal.

A different but related approach is taken in our work on SPI-Active (CoRL 2025): instead of adapting online during deployment, we do careful system identification before deployment using active exploration. By actively exciting the robot dynamics in simulation-informed ways, we identify the physical parameters of the specific real robot hardware (not just a nominal model) and build a calibrated simulator for policy training. This is adaptive identification in service of sim2real transfer rather than runtime adaptation.

Similarly, ASAP (RSS 2025) addresses the gap between simulation and real humanoid hardware by aligning the physics of the simulator to the real robot through parameter estimation — the same philosophy: treat the physical gap as something to estimate, not merely randomize over.

7. Open Questions

There is a lot of interesting open work at the intersection of adaptive control and robot learning. Here are the questions I find most compelling.

Adaptive control for hybrid systems. Classical adaptive theory was developed for smooth ODEs. Legged locomotion is fundamentally a hybrid dynamical system with discrete contact events. Extending Lyapunov-based stability guarantees to cover contact transitions — including the transient behavior during impact — is not fully solved. Some progress exists in hybrid adaptive control, but much remains to be done for the kinds of fast, dynamic gaits modern legged robots achieve.

Contact-consistent estimation. Adaptive estimators should explicitly respect the physical constraints imposed by contact — unilateral normal forces, friction cone membership — when estimating surface properties and object parameters. Standard least-squares estimators do not enforce these constraints, which can produce physically inconsistent estimates.

Meta-learning and fast adaptation. Can we combine the representational power of neural networks with the structure of adaptive control? Methods like MAML [2] and RMA [3] provide fast adaptation via learned update rules, but without the closed-loop stability guarantees of classical adaptive control. Bridging this gap — perhaps by learning an adaptation law that is provably stable — is an open problem.

When does learning beat adaptation? Not all uncertainty is low-dimensional and parametric. Visual uncertainty, terrain topology, and object shape are high-dimensional and may be better handled by learning. Characterizing which parts of the uncertainty should be handled by adaptive estimation (physics parameters) and which by learned representations (appearance, shape, semantics) is a useful framing for designing hybrid systems.

8. Conclusion

Adaptive control is not a relic of the pre-deep-learning era. It is a framework for handling structured, parametric physical uncertainty — and physical systems will always have this kind of uncertainty, regardless of how powerful our learned policies become. The key insight is that the sim2real gap has structure: it lives in a low-dimensional space of physical parameters embedded in known physical equations. This structure can be exploited by adaptive estimators in ways that domain randomization cannot.

For contact-rich systems — legged robots operating across varying terrain, dexterous hands manipulating novel objects — this matters most. Contact dynamics are the hardest to simulate accurately, and errors in contact models produce the most consequential failures at deployment time. Online estimation of contact and object properties is not just theoretically appealing; it is practically necessary for reliable deployment.

The most capable real-robot systems of the next few years will likely combine:

• Large-scale learned policies for behavior, perception, and high-level task structure;
• Adaptive control layers for online physical uncertainty estimation with principled guarantees;
• Careful sim2real pipelines that minimize the gap through system identification before deployment.

The next posts in this series will go deeper: a technical treatment of L1 adaptive control and its application to augmenting learned locomotion controllers, followed by a discussion of how online adaptation interacts with contact-rich manipulation.

References

1. N. Hovakimyan and C. Cao, L1 Adaptive Control Theory: Guaranteed Robustness with Fast Adaptation. SIAM, 2010.
2. C. Finn, P. Abbeel, and S. Levine, "Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks." ICML, 2017. [arXiv]
3. A. Kumar, Z. Fu, D. Pathak, and J. Malik, "RMA: Rapid Motor Adaptation for Legged Robots." RSS, 2021. [arXiv]
4. T. He et al., "ASAP: Aligning Simulation and Real-World Physics for Learning Agile Humanoid Whole-Body Skills." RSS, 2025. [arXiv]
5. N. Sobanbabu, G. He, T. He, Y. Yang, G. Shi, "Sampling-Based System Identification with Active Exploration for Legged Robot Sim2Real Learning." CoRL, 2025 (Oral). [arXiv]
6. Z. Xie, X. Da, B. Babich, A. Garg, and M. van de Panne, "Glide: Generalizable Quadrupedal Locomotion in Diverse Environments with a Centroidal Model." CoRL, 2023. (CaJun controller referenced above: [arXiv])