Contact-Anchored Policies (CAP)

• Contact-Anchored Policies (CAP) are a methodological paradigm that conditions policy actions on explicit physical contact information to improve precision and generalization.
• CAP integrates spatially grounded tactile, visual, and proprioceptive data to inform robust policy learning across robotic manipulation, locomotion, and network-based interventions.
• Empirical evaluations show CAP’s enhanced success rates and sample efficiency, demonstrating significant improvements over traditional methods in diverse applications.

Contact-Anchored Policies (CAP) constitute a methodological paradigm that conditions policy learning and execution—primarily in robotics and complex systems—on explicit representations of physical contact with the environment, rather than on abstract or indirect task descriptors. The central mechanism of CAP is the direct use of contact points, contact sequences, or spatially grounded tactile signals as input to policy models, enabling generalization, precision, and robustness in contact-rich or hybrid systems. This approach finds application in robot manipulation, locomotion, epidemiological intervention, and beyond, across a range of input modalities and computational architectures (Cui et al., 9 Feb 2026, Huang et al., 16 Oct 2025, Ciebielski et al., 2024, Zhao et al., 16 Jun 2025, Phillips-Grafflin et al., 2017, Fan et al., 2021, Zhang et al., 29 Jan 2026).

1. Formal Definitions and Foundational Principles

A Contact-Anchored Policy is defined by the explicit presence of contact-related information in the state or goal representation provided to the policy, anchoring decisions to physical points or moments of contact. In the canonical robotic manipulation setting, this takes the form of a policy

$\pi(a_t \mid s_t, c_t)$

where $s_t$ denotes the robot's sensory state and $c_t\in\mathbb{R}^3$ is a contact anchor, such as a 3D location in the robot's camera or hand frame corresponding to the site of intended or actual contact (Cui et al., 9 Feb 2026). CAP extends to multi-contact or temporally extended forms by conditioning on sequences of future or planned contacts, e.g.,

$g_{cc} = [p_{k+1},\, t_{rem}]$

with $p_{k+1}$ the next anticipated contact position and $t_{rem}$ the time until that contact (Ciebielski et al., 2024). In tactile policies, contact anchors are realized as spatially grounded features derived from tactile images registered to the robot's kinematic chain (Huang et al., 16 Oct 2025, Zhang et al., 29 Jan 2026).

The CAP paradigm generalizes beyond robotics. In epidemiological modeling, CAP refers to policies whose interventions are anchored on the empirical structure of observed contact networks, e.g., applying NPIs in response to measured contact intensity between individuals or groups (Fan et al., 2021).

2. Architectures and Mathematical Formulations

Contact-Anchored Policies are instantiated across a family of model architectures, unified by their explicit fusion of contact representations:

Policy Conditioning and Utility Modeling

• Robotic Manipulation: Policies are formulated as modular utility models $U(s,a;c)$, scoring actions $a$ given state $s$ and contact anchor $c$. This is implemented as a conditional imitation learner, minimizing

$$\mathcal{L} = -\sum_{t=1}^T \log \pi(a_t | o_{t-k:t},\, p_{t-k:t})$$

where $(o_t, p_t)$ are temporal sequences of sensory observations and inferred or planned contact positions (Cui et al., 9 Feb 2026).

• Locomotion: For multi-gait control, CAP represents goals by upcoming contact events.
  • Single-contact conditioning: $g_{cc} = [p_{k+1}, t_{rem}]$
  • Two-contact conditioning: $$g_{tcc} = [p_{k+1}, t_{rem}, p_{k+2}, d_{nom} + t_{rem}]$$
  • Policies map joint state and contact goal to actions:

$\pi: S \times G \to A$

Training is performed via behavioral cloning from MPC-generated data (Ciebielski et al., 2024).

• Visuo-Tactile Skill Policies: Tactile features $T_t$ and vision $I_t$ are encoded and fused, typically as concatenated latent tokens. FiLM (Feature-wise Linear Modulation) layers, forward kinematics-based spatial anchoring, and transformer-based sequence modeling are used to maintain alignment between tactile events and the robot's body in space (Huang et al., 16 Oct 2025, Zhao et al., 16 Jun 2025, Zhang et al., 29 Jan 2026).

Spatial Anchoring Mechanics

A distinguishing property of CAP is the anchoring of observations or outputs in a specific spatial frame:

• Hand/frame anchoring: Contact features (e.g., tactile pixels or center-of-pressure) are mapped into the hand or end-effector's coordinate system using the known kinematics $T_{hand}^{sensor}(q)$, usually via:

$p_h = T_{hand}^{sensor}(q) \cdot p_s$

• Contact event fusion: Contact switch times, tactile events, or spatial contact points are concatenated with proprioceptive and visual features to anchor the policy in space and time (Huang et al., 16 Oct 2025, Ciebielski et al., 2024).
• Belief-space planning: In contact-uncertain environments, CAP uses particle-based sampling to plan over possible contact outcomes, splitting the belief tree at contact events and updating action selection accordingly (Phillips-Grafflin et al., 2017).

3. Data Collection, Training, and Optimization Procedures

CAP leverages data- and simulation-driven pipelines, emphasizing sample efficiency relative to high-parameter, language-conditioned models.

• Demonstration Gathering: Contact-rich demonstrations are acquired by kinesthetic teaching (Zhang et al., 29 Jan 2026), VR teleoperation (Zhao et al., 16 Jun 2025), or handheld gripper data collection (Cui et al., 9 Feb 2026). Contact events are labeled in hindsight, often by monitoring gripper closure or tactile event activation.
• Feature Preprocessing: Contact positions, tactile detection events, center-of-pressure vectors, and force distributions are precomputed, enabling policies to attend to both the “where” and “how” of contact (Zhang et al., 29 Jan 2026). Forward kinematics is used to register tactile features to the kinematic frame (Huang et al., 16 Oct 2025).
• Chunked Policy Outputs: Network outputs often take the form of action chunks (e.g., 10 or 100 future actions) for temporal smoothness (Zhao et al., 16 Jun 2025, Huang et al., 16 Oct 2025).
• Objective Functions: Imitation learning dominates, with mean squared error, negative log likelihood, or chunked action-matching as loss terms. Latent variable regularization is sometimes added (Zhang et al., 29 Jan 2026, Huang et al., 16 Oct 2025).

4. Empirical Evaluation and Quantitative Impact

CAP has demonstrated state-of-the-art outcomes in a range of real-world and simulated tasks:

System/Task	Baseline	CAP Variant	Success Rate Gain	Notes
Manipulation (Pick)	VLA, AnyGrasp (47%)	CAP (83%)	+36%	Zero-shot, novel objects, Oracle anchor
Loco-manipulation (CAP₁)	Velocity Cond. (VC)	CAP₁	2-3× lower fail	Gait switches, 30% lower lateral error
Tactile Dexterous (SaTA)	Flat/Global Tactile	SaTA	+30pp (SR), -27% time	Sub-mm alignment, FC: +25–45pp
Visuo-tactile (ViTaL)	ViSk, BAKU (~40-50%)	ViTaL (~90%)	+40pp	Rich tactile anchoring, distractors robust
High-precision (DexTac)	Force-only (60%)	DexTac (92%)	+32%	Sub-mm injections, CoP anchoring crucial
Pandemic Policy (COVID-19)	Unanchored NPIs	CAP (Network-anchored)	Non-dominated Pareto front	Multi-county application

Empirically, CAP reduces the required demonstration data and parameter count while preserving or improving performance relative to end-to-end VLAs, particularly in generalization (unseen objects, environments, or robot embodiments) (Cui et al., 9 Feb 2026, Zhang et al., 29 Jan 2026).

5. Design Insights, Failure Modes, and Limitations

Contact anchoring introduces several key advantages and necessitates specific design considerations:

• Shared structure: Policies anchored on contacts exploit shared structure across tasks or gaits, yielding smoother mode transitions (e.g., walking-to-running) and reducing the burden of inferring dynamic transitions from distal task cues (Ciebielski et al., 2024).
• Spatial Generalization: Anchoring in body frames or end-effector coordinates supports robust transfer across scene layouts and robot morphologies (Huang et al., 16 Oct 2025, Cui et al., 9 Feb 2026).
• Tactile Diversity: Multi-dimensional tactile encoding (force, CoP, image features) is essential for stable contact maintenance; loss of spatial grounding leads to drift, slip, or alignment failures (Huang et al., 16 Oct 2025, Zhang et al., 29 Jan 2026).
• Simplicity vs. Complexity: Single-contact anchoring suffices for many tasks; two-contact heuristics may overfit or be brittle when disturbance breaks expected contact order (Ciebielski et al., 2024).
• Failure Modes: CAP models can fail on ambiguous or incorrectly specified contact prompts; end-to-end reliability can be increased by integrating verifier-retry mechanisms or shifting to multi-anchor models (Cui et al., 9 Feb 2026).
• Limitations: Current CAP architectures typically address single-contact or atomic actions; generalizing to complex sequences or bimanual tasks is an open area.

6. Applications Beyond Manipulation: Epidemiology and Hybrid Systems

Beyond robotics, CAP formalism applies to systems where policies must target, modulate, or adapt interventions based on the structure of a contact network:

• In COVID-19 containment, CAP identifies Pareto-optimal mixes of NPIs (masking, mobility restrictions, reopening) anchored to observed contact intensity in the agent interaction network. The modeled outcome achieves locally optimal trade-offs between economic cost and infection count, demonstrating the generality of CAP as a network-driven policy design principle (Fan et al., 2021).
• In manipulation under uncertainty, belief-space CAPs enable resilient execution via real-time adaptation to unexpected contact events, leveraging sampling-based planners and online policy graph updates (Phillips-Grafflin et al., 2017).

7. Implications and Outlook

Contact-Anchored Policies represent a shift from abstract, language-based, or globally indexed task specification to a paradigm where the physical geometry of the task—manifest as explicit spatial or temporal contact representations—is integral to policy input. This suggests that robust generalization in contact-rich environments, data efficiency, and system modularity can be addressed by aligning action selection as tightly as possible with the physical substrate of interaction. Open research directions include extension to multi-contact, long-horizon reasoning, integration with high-level planners or verifiers, and translation into further domains where interaction networks govern system evolution (Cui et al., 9 Feb 2026, Fan et al., 2021).