---
title: Contact Residual Dynamics
url: https://www.emergentmind.com/topics/contact-residual-dynamics-crd
type: topic
---

# Contact Residual Dynamics

Searching arXiv for recent papers on Contact Residual Dynamics and closely related residual contact modeling.
Tool call: arxiv_search(query="Contact Residual Dynamics MPC learned contact residuals thruster locomotion residual contact model", max_results=10, sort_by="relevance")
Contact Residual Dynamics (CRD) denotes a class of residual modeling methods for contact-rich robotic systems in which discrepancies between a nominal or prior model and actual contact behavior are learned and then injected back into prediction, estimation, or control. In the cited literature, CRD appears in several technically distinct forms: as learned residual angular dynamics caused by leg-ground impacts in thruster-assisted quadrupedal locomotion [2508.03003], as an additive residual in contact complementarity constraints for adaptive contact-implicit MPC [2310.09893], as residual corrections to analytical contact impulses for long-horizon prediction and uncertainty propagation [2009.03994], and, in a broader sensory formulation, as residual tactile representations that encode the discrepancy between visual priors and physical contact sensations [2607.03387]. A later line of work further situates CRD within stochastic residual modeling by arguing that contact residuals are often multi-modal rather than merely poorly modeled, and therefore benefit from structured stochastic representations [2603.08478].

## 1. Conceptual scope and definitions

CRD is fundamentally a residual-dynamics formulation for contact phenomena that are difficult to capture with nominal rigid-body models, simplified actuated models, or visually predicted interaction priors. The recurring technical pattern is decomposition: a structured nominal model captures the dominant predictable component, while a learned residual captures unmodeled interaction effects such as impacts, frictional variability, hybrid mode mismatch, or visually unpredictable tactile events [2508.03003].

In thruster-enhanced locomotion, the residual is defined at the level of body angular acceleration. The true angular acceleration is written as the sum of the nominal effect of thrusters and residual effects due to leg-ground impact [2508.03003]:
\[
\dot{\boldsymbol{\omega}}_{\text{true}}
=
\dot{\boldsymbol{\omega}}_{\text{nominal}}
+
\dot{\boldsymbol{\omega}}_{\text{residual}}.
\]
There, CRD specifically approximates impulsive torques from ground contacts that affect the robot’s orientation, especially roll [2508.03003].

In adaptive contact-implicit MPC, CRD is defined differently. The residual is inserted additively into the contact complementarity constraints of a local Linear Complementarity System (LCS), where the vector \(r_{\text{comp}}\) represents model discrepancies or errors in the contact complementarity constraints and is learned online to minimize predictive error [2310.09893]. In that formulation, CRD primarily adapts the hybrid model boundaries, meaning the transitions between contact and no-contact modes [2310.09893].

In residual contact simulation, the residual is a corrective impulse added to an analytical contact impulse model. Given a nominal analytical impulse \(\mathbf{p}_m\), the learned contact residual is \(\mathbf{p}_{res}\), yielding
\[
\mathbf{p}_{opt} = \mathbf{p}_m + \mathbf{p}_{res},
\]
with the post-contact state updated by the corrected impulse [2009.03994]. This places CRD at the interface between analytical simulators and empirical data.

A broader interpretation appears in contact-rich manipulation with multimodal sensing. There, tactile information is reformulated as a residual quantity,
\[
r_t = z_t - \hat{z}_t,
\]
where the residual captures the discrepancy between visually predicted tactile latent state and actual tactile perception [2607.03387]. This is presented as a residual representation of unexpected contact information rather than a force or state residual in a dynamics equation.

Taken together, these uses indicate that CRD is not a single fixed model class. It is a family of residual formulations centered on contact-induced discrepancy, with the residual placed at different levels of the pipeline: continuous dynamics, hybrid constraints, contact impulses, or cross-modal sensory representation.

## 2. Mathematical formulations across representative CRD systems

The most explicit dynamics-based CRD formulation in locomotion appears in the Husky-\(\beta\) system, where nominal angular acceleration comes from thruster torques [2508.03003]:
\[
I\,\dot{\boldsymbol{\omega}}_{\text{nominal}}
=
\sum_{i=1}^4
\mathbf{r}_i \times
\left(c_f\, v_i^2\, \mathbf{\hat{e}}_i\right).
\]
The residual component approximates the effect of ground reaction forces at the legs,
\[
\dot{\boldsymbol{\omega}}_{\text{residual}}
\approx
I^{-1}
\sum_{i=1}^4
\left(\mathbf{d}_i \times \mathbf{F}_i\right),
\]
and the learned version becomes
\[
\dot{\boldsymbol{\omega}}_{\text{residual}}
=
I^{-1}
\sum_{i=1}^4
C_{i,\text{NN}}
\left(
\mathbf{d}_i \times \mathbf{F}_{i,\text{NN}}
\right),
\]
where each leg contributes a predicted ground reaction force weighted by a predicted contact probability [2508.03003]. The resulting augmented thrust model is
\[
\begin{aligned}
\dot{\boldsymbol{\theta}} &= \boldsymbol{\omega}, \\
\dot{\boldsymbol{\omega}} &=
I^{-1}
\sum_{i=1}^{4}
\left[
\mathbf{r}_i \times (c_f v_i^2 \mathbf{\hat{e}}_i)
+
C_{i,\text{NN}} (\mathbf{d}_i \times \mathbf{F}_{i,\text{NN}})
\right].
\end{aligned}
\]
This formulation preserves a physically structured nominal term and adds a contact residual torque term inside the predicted angular dynamics [2508.03003].

In adaptive hybrid MPC, the prior multi-contact model is represented as an LCS,
\[
\begin{aligned}
x_{k+1} &= A^* x_k + B^* u_k + D^* \lambda_k + d^*, \\
0 &\leq \lambda_k \perp E^* x_k + F^* \lambda_k + H^* u_k + c^* \geq 0,
\end{aligned}
\]
and CRD modifies the complementarity relation to
\[
\begin{aligned}
x_{k+1} &= A^* x_k + B^* u_k + D^* \lambda_k + d^*, \\
0 &\leq \lambda_k \perp E^* x_k + F^* \lambda_k + H^* u_k + c^* + r_{\text{comp}} \geq 0.
\end{aligned}
\]
In this case, the residual is not primarily a force correction but a correction to the hybrid contact law itself [2310.09893].

In residual point-contact learning, the analytical model produces
\[
\mathbf{p}_m = \mathbf{m}(\mathbf{s}^{pre}, \mu, \epsilon, \mathbf{M}, \mathbf{f}_{ext}),
\]
and the contact-corrected post-impact velocity is
\[
\mathbf{v}^{post}
=
\mathbf{v}^{pre}
+
\mathbf{M}^{-1}
(\mathbf{f}_{ext} dt + \mathbf{J}^T (\mathbf{p}_m + \mathbf{p}_{res})).
\]
This formulation is localized to contact events and emphasizes residual learning at the impulse level rather than along continuous dynamics [2009.03994].

A later extension argues that deterministic contact residuals are insufficient when contact outcomes are multi-modal. STRIDE decomposes predicted acceleration into a deterministic Lagrangian component and a stochastic residual force sample [2603.08478]:
\[
\ddot{\mathbf{q}}_{\text{pred}}
=
f_{\text{LNN}}(\mathbf{q}, \dot{\mathbf{q}}, \boldsymbol{\tau}; \theta)
+
\mathbf{M}^{-1}(\mathbf{q})\;
\boldsymbol{\epsilon}_{\text{CFM}}(\mathbf{q}, \dot{\mathbf{q}}, \boldsymbol{\tau}, \mathbf{z}; \phi).
\]
This work explicitly states that traditional CRD associates unmodeled effects such as contacts and friction with deterministic, state-dependent residuals, and presents stochastic residuals as an extension that avoids averaging bias [2603.08478].

## 3. Learning architectures and training objectives

The CRD learning architecture in thruster-enhanced locomotion uses one subnetwork per leg, with four identical neural nets in total [2508.03003]. Each subnetwork processes a 21D feature vector comprising joint angles and velocities, foot and propeller positions, base orientation, angular and linear velocities, and thruster actions [2508.03003]. Each network outputs a predicted 3D ground reaction force \(\mathbf{F}_{i,\text{NN}}\) and a predicted contact probability \(C_{i,\text{NN}} \in [0,1]\), and the effective force per leg is \(\mathbf{F}_{i,\text{NN}} \cdot C_{i,\text{NN}}\) [2508.03003].

Its training objective is physics-informed. The force loss is
\[
L_{\text{GRF}}
=
\left\|
I^{-1}\sum_{i=1}^4
C_{i,\text{NN}}
\left( \mathbf{d}_i \times \mathbf{F}_{i,\text{NN}} \right)
-
(\dot{\boldsymbol{\omega}}_{\text{true}} - \dot{\boldsymbol{\omega}}_{\text{nominal}})
\right\|^2,
\]
the contact loss is binary cross-entropy,
\[
\begin{aligned}
L_{\text{contact}}
=
-\frac{1}{4} \sum_{i=1}^4
\left[
C_{i,\text{GT}} \log C_{i,\text{NN}}
+
(1 - C_{i,\text{GT}}) \log (1 - C_{i,\text{NN}})
\right],
\end{aligned}
\]
and the overall loss is
\[
L = (1-\alpha) L_{\text{GRF}} + \alpha L_{\text{contact}}.
\]
The model is trained offline with data from simulation and/or real hardware, and when no contact sensors are available on hardware, simulation-trained weights are used with the contact head frozen [2508.03003].

In adaptive contact-implicit MPC, CRD is learned online rather than offline. The pipeline collects tuples \((x_k, u_k, x_{k+1})\), recomputes local LCS matrices at each operating point, defines a batch loss measuring mismatch between predictions and observations, and updates \(r_{\text{comp}}\) using gradient descent or Adam [2310.09893]. The loss has a state-prediction term and a complementarity-violation term:
\[
L_\epsilon (\mathcal{B}^{A}, r_{\text{comp}}) = \sum_{k} l_\epsilon(x_{k+1}, x_k, u_k, \theta^*_k, r_{\text{comp}}),
\]
with each sample loss involving minimization over nonnegative contact force and slack variables [2310.09893]. The paper states that the gradient with respect to \(r_{\text{comp}}\) is computed using implicit differentiation and that residual learning runs at approximately \(20\) Hz while MPC runs at approximately \(80\) Hz [2310.09893].

The residual point-contact learner adopts a density neural network that maps contact features to a distribution over corrective impulses [2009.03994]. The feature vector uses effective inertia at contact and contact point velocity rather than the full state, yielding a 5-dimensional planar feature representation [2009.03994]. The model is stochastic, with the residual impulse sampled from a Gaussian whose mean and covariance depend on the input features. Training minimizes the expected trajectory discrepancy between observed and estimated trajectories and uses gradient-free optimization because contact timing is noisy and semi-Markovian [2009.03994].

STRIDE’s training objective differs by jointly learning a structured Lagrangian prior and a stochastic residual process. The Conditional Flow Matching module generates a residual by integrating a conditional ODE vector field from latent Gaussian noise [2603.08478]:
\[
\boldsymbol{\epsilon}_{\text{CFM}}
=
\mathbf{z}_0 + \int_0^1 v_\phi(\mathbf{z}_t, t \mid \mathbf{c})\, dt,
\qquad
\mathbf{z}_0 \sim \mathcal{N}(\mathbf{0}, \mathbf{I}),
\]
and training minimizes a supervised regression objective on accelerations [2603.08478]. The paper explicitly frames this as an extension beyond deterministic CRD.

## 4. Integration into model predictive control and predictive systems

A central role of CRD is to improve model-plant alignment inside predictive controllers. In the Husky-\(\beta\) locomotion architecture, control is decoupled because a unified MPC optimizing both ground reaction forces and thruster forces is limited by the low torque-control bandwidth of lightweight actuators [2508.03003]. The legged controller is Raibert-type and position-based, while an MPC regulates the thrusters using a simplified dynamics model augmented with CRD [2508.03003]. Dynamics are linearized and discretized as
\[
\mathbf{x}_{k+1} = A_d \mathbf{x}_k + B_d \mathbf{u}_k,
\]
and the CRD output is incorporated into the reference state used over the MPC horizon:
\[
\mathbf{x}_{r,k}
=
\begin{bmatrix}
\boldsymbol{\theta}_0 + k \Delta t (\boldsymbol{\omega}_d + k \Delta t \dot{\boldsymbol{\omega}}_{\text{residual}}) \\
\boldsymbol{\omega}_d + k \Delta t \dot{\boldsymbol{\omega}}_{\text{residual}}
\end{bmatrix}.
\]
The MPC then solves
\[
\begin{aligned}
\min_{\mathbf{u}} &\qquad (\mathbf{x} - \mathbf{x}_r)^\top \bar{Q} (\mathbf{x} - \mathbf{x}_r) + \mathbf{u}^\top \bar{R} \mathbf{u} \\
\text{s.t.} &\qquad 0 < u_{t,i} < u_t^{\text{max}},\quad i=1,\ldots,4.
\end{aligned}
\]
Because \(\mathbf{x}_r\) includes CRD-informed roll, pitch, and yaw trajectories, the reference better aligns prediction with actual dynamics during ground-contact events [2508.03003].

In adaptive contact-implicit MPC, integration is more direct: the MPC always uses the latest updated residual-augmented LCS model \((\theta^*, r_{\text{comp}})\) to compute control inputs [2310.09893]. This allows the controller to adapt the hybrid model boundary online, rather than only correcting continuous dynamics [2310.09893]. The framework thereby couples online system identification of contact discrepancies with real-time contact-implicit control.

Outside MPC, CRD also enters long-horizon prediction pipelines. In residual point-contact learning, the residual is applied only when a collision occurs; between contacts, free motion is propagated deterministically [2009.03994]. This yields a contact-aware predictive simulator in which uncertainty is injected at contact events and then propagated through state belief updates [2009.03994]. A plausible implication is that this event-triggered residual placement can reduce unnecessary correction outside contact phases, which is consistent with the paper’s emphasis on sample efficiency and physically structured corrections [2009.03994].

## 5. Empirical behavior, performance, and operational significance

In thruster-enhanced locomotion, offline test predictions show that the CRD-augmented model achieves lower RMSE in angular acceleration \(\dot{\boldsymbol{\omega}}\), especially for roll, than a nominal-only model [2508.03003]. In push-recovery experiments with a \(15\)N disturbance applied for \(0.5\)s along the \(y\)-axis, the CRD-augmented controller successfully recovers to a stable trot, whereas the nominal-only controller fails to recover stability [2508.03003]. In hardware narrow-path cat gait trials, the CRD-augmented controller achieves lower RMSE in roll, pitch, and yaw than the nominal dynamics alone, with the largest improvements in roll stability [2508.03003]. In standard trot gait, it maintains stable roll with less required thruster force than the nominal-only case, which the paper interprets as better alignment with the true system dynamics [2508.03003].

In adaptive multi-contact manipulation, the online residual learning framework is reported to adapt on-the-fly with an adaptation rate around \(20\) Hz and to run MPC around \(80\) Hz [2310.09893]. Hardware experiments show that with a rough prior model, the framework successfully manipulates previously unknown objects with non-smooth surface geometries [2310.09893]. The reported qualitative outcome is that model-based control without adaptation fails, whereas the adaptive CRD-enhanced version enables required contact behavior on hardware [2310.09893].

In long-horizon residual point-contact learning, the proposed point-contact residual model converges to about \(25\%\)–\(30\%\) lower RMSE over full trajectories than the bare simulator, and does so with as few as \(5\)–\(10\) trajectories [2009.03994]. The paper also reports that a generic recurrent residual baseline, SIAN, requires the entire dataset before matching bare simulator performance [2009.03994]. The same work emphasizes uncertainty “blossoming” after each contact and uses Gaussian-mixture belief propagation to represent the resulting multi-modality [2009.03994].

STRIDE reports a \(20\%\) reduction in long-horizon prediction error and a \(30\%\) reduction in contact force prediction error compared to deterministic residual baselines [2603.08478]. On the Unitree Go1 benchmark summarized in the paper, STRIDE attains state RMSE \(0.932\) and force error \(6.7\), compared with \(1.326\) and \(9.4\) for DeLaN and \(1.154\) and \(8.8\) for LNN + Diffusion [2603.08478]. The paper attributes these gains to physically structured modeling and stochastic residual representation of contact phenomena [2603.08478].

In tactile manipulation, ResTacVLA reports \(62.8\%\) average success on a diverse task suite, compared with \(42.3\%\) for the best baseline and \(28.2\%\) for a vision-only model [2607.03387]. The paper further reports that removing vector quantization reduces average performance by \(26.7\%\), and removing surprise-aware gating reduces it by \(13.3\%\) [2607.03387]. Although this is not a dynamics-equation formulation of CRD, it shows that residual contact representations can also improve robustness to unexpected dynamic disturbances in action models [2607.03387].

## 6. Deterministic versus stochastic CRD, and related interpretations

A significant conceptual development in the literature is the distinction between deterministic residual correction and stochastic residual modeling. In locomotion CRD and online hybrid MPC, the residual is deterministic once conditioned on state, action, and learned parameters [2508.03003; 2310.09893]. In residual point-contact learning, the residual is explicitly stochastic and represented as a Gaussian correction at contact events, with uncertainty propagated through the subsequent state [2009.03994]. STRIDE makes this distinction explicit and argues that deterministic residual methods average over incompatible outcomes in multi-modal settings such as slipping versus sticking [2603.08478]. The paper calls this limitation “averaging bias” and presents stochastic residuals as a remedy [2603.08478].

This produces two broad interpretations of CRD. The first treats CRD as a deterministic, state-dependent correction to a nominal model. The second treats CRD as an uncertain or multi-modal interaction term whose distribution must itself be modeled. The available papers support both views. A plausible synthesis is that deterministic CRD is adequate when contact discrepancies are repeatable and sufficiently observable, whereas stochastic CRD becomes necessary when unobserved micro-conditions make contact outcomes inherently variable [2603.08478; 2009.03994].

Another interpretive distinction concerns where the residual is inserted. Force-level CRD modifies torques, impulses, or generalized forces [2508.03003; 2009.03994; 2603.08478]. Constraint-level CRD modifies complementarity relations and therefore contact mode transitions [2310.09893]. Representation-level CRD encodes the mismatch between sensory expectation and sensed contact [2607.03387]. This suggests that CRD is best understood as a residual-contact principle rather than a single canonical equation.

## 7. Relationship to adjacent research directions and open issues

CRD is closely related to residual physics, data-augmented simulation, contact-implicit control, and multimodal contact representation. The residual point-contact framework explicitly combines an analytical contact simulator with learned corrections, emphasizing that purely analytical models cannot match observed diversity of outcomes while generic data-driven models require substantially more data [2009.03994]. Adaptive contact-implicit MPC extends residual learning into online hybrid control, addressing the difficulty of acquiring accurate multi-contact models without extensive offline tuning [2310.09893]. Thruster-enhanced locomotion uses CRD to compensate for the infeasibility of unified torque-level MPC under actuator bandwidth limitations [2508.03003]. STRIDE, in turn, connects CRD to structured mechanics and generative stochastic modeling [2603.08478].

The literature also reveals several recurring technical challenges. One is hardware observability: in Husky-\(\beta\), transfer learning is needed when no contact sensors are available on hardware, and the contact head is frozen during transfer [2508.03003]. Another is real-time optimization: adaptive hybrid MPC requires concurrent residual learning and control loops, with residual updates at around \(20\) Hz and MPC at around \(80\) Hz [2310.09893]. A third is contact multi-modality and uncertainty propagation, which motivates Gaussian or flow-based residual models in the simulation and structured stochastic modeling literature [2009.03994; 2603.08478].

A common misconception is to equate CRD exclusively with force estimation. The surveyed formulations show that CRD may estimate forces, but it may also correct hybrid constraints, refine analytical impulses, or encode cross-modal contact discrepancy [2310.09893; 2009.03994; 2607.03387]. Another misconception is that residuals merely patch poor models in an ad hoc manner. The strongest examples in this literature retain explicit physical structure: nominal thruster dynamics plus residual torque [2508.03003], LCS contact laws plus residual complementarity term [2310.09893], or Lagrangian mechanics plus stochastic residual interactions [2603.08478].

The current evidence suggests that CRD is most effective when it is physically situated: tied to contact geometry, contact timing, or hybrid contact constraints rather than learned as an unrestricted black-box correction. A plausible implication is that future CRD systems will continue to combine structured priors with residual terms that are event-aware, uncertainty-aware, and suitable for direct insertion into planners or controllers. The available papers already point in that direction through physics-informed residual losses [2508.03003], online residual adaptation inside contact-implicit MPC [2310.09893], uncertainty-aware contact correction [2009.03994], and stochastic structured residual dynamics [2603.08478].

Source: https://www.emergentmind.com/topics/contact-residual-dynamics-crd