Contact Trust Region (CTR) Framework

• The Contact Trust Region (CTR) framework is a suite of algorithmic and geometric methods that explicitly incorporate unilateral and non-smooth contact constraints for advanced optimization and decision-making.
• It combines local Taylor- or mirror-map approximations with specialized trust regions to maintain feasibility, supporting applications in dexterous robotic manipulation, nonlinear elasticity, and constrained reinforcement learning.
• Empirical results show CTR’s ability to achieve low tracking errors and efficient convergence in both online MPC and offline roadmap-based planning, outperforming standard methods in contact-rich scenarios.

The Contact Trust Region (CTR) framework encompasses a suite of algorithmic and geometric tools for solving contact-rich decision and optimization problems. CTR methodologies generalize classical trust region techniques by explicitly incorporating the unilateral and often non-smooth constraints that arise in physical and decision-theoretic contact. Applications span dexterous robotic manipulation, nonlinear elasticity with contact, and safety-constrained reinforcement learning. Distinct variants include the contact-aware trust region for convex trajectory optimization in manipulation (Suh et al., 4 May 2025), filter–trust-region methods for large deformation mechanics (Youett et al., 2017), and constraint-barrier trust region policy updates in reinforcement learning (Milosevic et al., 2024). CTR techniques combine local Taylor- or mirror-map-based modeling with contact-specific feasibility domains, resulting in scalable, globally convergent solutions.

1. Contact Trust Region in Dexterous Manipulation

The Contact Trust Region as developed in "Dexterous Contact-Rich Manipulation via the Contact Trust Region" (Suh et al., 4 May 2025) provides a principled local approximation for contact dynamics in manipulation tasks. The underlying model is quasidynamic, with the configuration $q \in \mathbb{R}^{n_q}$ governed by an SOCP, capturing both actuated ("robot") and unactuated ("object") coordinates. At each decision step, the system evolves according to:

• $q_+ = f(q, u)$, where $u$ is the commanded robot position.
• Evolution is subject to velocity cone constraints $J_i(q) q_+ + c_i(q) \in \mathcal{K}_i$ enforcing nonpenetration and friction.

Rather than employing a standard ellipsoidal trust region, CTR defines the region of trust as the set of $(\delta q, \delta u)$ for which first-order Taylor approximations of the next state $q_+$ and contact force $\lambda$ are guaranteed to satisfy both primal (no penetration) and dual (friction cone) constraints:

• $$\mathcal{S}_{\Sigma,\kappa}(\bar{q}, \bar{u}) = \{ (\delta q, \delta u) \in \mathcal{E}_\Sigma \mid$$ linearized $q_+$ and $\lambda_+$ feasible for all contacts$q_+ = f(q, u)$0.

Two variants are distinguished: strict CTR, enforcing both primal and dual, and Relaxed CTR (R-CTR), relaxing the nonpenetration for numerical performance but maintaining dual (force) feasibility in the linearized contact cone.

2. CTR-Based Local Model Predictive Control (MPC) Formulation

CTR is integrated into a local, finite-horizon MPC scheme by solving a convex trajectory optimization problem with dynamics linearized around a nominal trajectory. The optimization:

• Minimizes terminal deviation and control smoothness: $q_+ = f(q, u)$1.
• Subjects the first-order dynamics to the linearized contact model and R-CTR constraints at every time step.
• The resulting subproblem is a sequence of SOCP stages that enforce dual-cone feasibility (or both primal and dual in the strict variant).

Typically, 1–3 sequential convexification steps suffice to converge, thanks to the stability imparted by the contact-aware trust region.

3. Algorithmic Realization of CTR-MPC

CTR-MPC algorithms involve two principal routines:

• Offline trajectory optimization (CtrTrajOpt): Iteratively roll out the nonlinear contact model, solve the convex trust-region subproblem, and update the nominal trajectory.
• Online MPC loop: At each step, warm-start the trajectory and solve for an optimized local control sequence, applying only the first control, then advancing the system and repeating.

An initial-guess heuristic, wherein a virtual torque is applied to establish contact, accelerates convergence when the system starts out-of-contact.

4. Global Planning via Roadmaps

While CTR-MPC provides high-fidelity local planning, it is complemented by a roadmap-based global strategy for contact-rich systems with complex mode transitions:

• Construct an offline graph whose nodes correspond to stable object-robot contact configurations.
• Edges encode feasible local transitions, determined by successful runs of CTR-MPC and collision-free arm repositioning.
• Online, connect the start and goal to nearest roadmap nodes using MPC, then execute the path yielded by shortest-path search.

Roadmap construction for a high-DOF dexterous hand (e.g., AllegroHand, covering all 24 symmetries across 5 grasps and ≈100 edges) can be accomplished in under 10 minutes using CPU resources.

5. Empirical Performance in Manipulation

CTR-based planners have been evaluated both in simulation and hardware for planar (IiwaBimanual) and 3D (AllegroHand) systems:

• Achieved local tracking errors of $q_+ = f(q, u)$2 mm/2.1 mrad (R-CTR) and $q_+ = f(q, u)$3 mm/8.9 mrad (CTR) for IiwaBimanual, $q_+ = f(q, u)$4 mm/$q_+ = f(q, u)$5 mrad for AllegroHand.
• R-CTR outperforms both standard ellipsoidal trust region baselines and strict CTR in mean/variance.
• Per MPC iteration wall-clock: $q_+ = f(q, u)$6 ms (R-CTR, IiwaBimanual) to $q_+ = f(q, u)$7 ms (CTR, AllegroHand).
• Roadmap-based global plans are constructed offline in $q_+ = f(q, u)$8 minutes, with online execution at several Hz using only CPU.
• Compared to RL-based approaches requiring thousands of GPU-hours for training, CTR offers comparable hardware dexterity with orders-of-magnitude lower compute (Suh et al., 4 May 2025).

6. Filter–Trust-Region Methods in Hyperelastic Contact

In numerical mechanics, the filter–trust-region framework (Youett et al., 2017) solves large deformation contact problems by augmenting the quadratic model with a filter that balances energy decrease and infeasibility reduction. Contact is imposed via mortar discretization, which ensures numerical stability and avoids spurious oscillations. The trust-region subproblem is a QP with $q_+ = f(q, u)$9 bounds, rapidly solved using Truncated Nonsmooth Newton Multigrid (TNNMG) with Monotone Multigrid correction for nonconvexities. Global convergence is guaranteed by filter-acceptance criteria and feasibility-restoration phases.

7. Constrained Trust Region Policy Optimization in Reinforcement Learning

In constrained reinforcement learning, the C-TRPO algorithm (Milosevic et al., 2024) introduces a trust region whose shape is governed by a composite mirror map including barrier terms for safety constraints:

• The trust region consists only of policies strictly within the safety set, as the Bregman divergence blows up at constraint boundaries.
• At each policy update, a quadratic subproblem is solved with the reward gradient subject to the barrier-augmented trust region.
• The result is monotonic improvement in reward and drastic reduction in cumulative constraint violation compared to prior safe RL algorithms, while preserving computational efficiency.

C-TRPO recovers standard TRPO and Constrained Policy Optimization as special cases, offering safety invariance and global convergence to optimal safe policies.

---

References:

• Contact-rich manipulation: (Suh et al., 4 May 2025)
• Large deformation contact mechanics: (Youett et al., 2017)
• Constrained RL and information geometry: (Milosevic et al., 2024)