---
title: Contact-Explicit Policy Learning
url: https://www.emergentmind.com/topics/contact-explicit-policy-learning
type: topic
---

# Contact-Explicit Policy Learning

Searching arXiv for recent papers on contact-explicit policy learning and closely related force/tactile-conditioned methods.
Contact-explicit policy learning denotes a class of robot-learning methods that treat contact as a first-class computational object rather than an unmodeled consequence of geometry or kinematics. In recent work, contact is represented directly as external torque estimates, tactile contact states, objective contact positions and binary flags, contact event sequences, part-level contact commands, or contact-induced motion fields, and those representations are then used to condition policies, reshape training distributions, define rewards, refine simulators, or organize hybrid controllers [2606.12406] [2601.12796] [2411.12982] [2510.03599] [2607.08742] [2606.30268].

## 1. Representational forms of explicit contact

The literature uses several distinct contact representations, each tied to a different control or learning objective.

| Formulation | Explicit variable | Primary role |
|---|---|---|
| FACTR 2 [2606.12406] | $\hat{\tau}_{ext}$ and $s_t \in \{F, PC, C\}$ | policy input and phase-aware re-sampling |
| HDP [2411.12982] | objective contact $C_t$ | high-level contact guidance for low-level diffusion actions |
| Unified contact-goal RL [2510.03599] | $p^{con}_{t,e}$, $I^{con}_{t,e}$, $S$, $s$ | goal-conditioned multi-task locomotion/manipulation |
| ContactMimic [2607.08742] | per-part binary contact commands $c_t$ | runtime contact controllability |
| ConCent [2606.30268] | contact event sequence $E=\{e_s\}$ | simulator grounding and stage-gated reward shaping |

In FACTR 2, contact is explicit at two levels: as an estimated external torque signal $\hat{\tau}_{ext}$ and as phase labels derived from it, with timesteps partitioned into Free Space, Pre-Contact, and Contact [2606.12406]. In "Contact-Aware Neural Dynamics" [2601.12796], contact is a binary or probabilistic signal $c_t$, predicted into the future and fused into a contact-conditioned residual dynamics model. In PhaForce, the explicit schedule consists of a continuous contact probability $p_t^c$ and a soft phase belief vector $p_t$, which jointly determine when force enters planning and where corrective residuals are routed [2603.08342].

Other formulations elevate contact from a sensed variable to a task specification. The multi-task framework in "Learning to Act Through Contact" [2510.03599] defines tasks through sequences of desired contact positions, timings, and active end-effectors. HDP predicts objective contacts $C_t$ as per-finger 3D positions with binary contact flags and uses them to guide a low-level diffusion actor [2411.12982]. ContactMimic conditions a humanoid policy on per-part binary contact commands, so the same geometry can be executed with contact either produced or suppressed [2607.08742]. ConCent treats the demonstrated contact event sequence as the central real-to-sim-to-real learning signal [2606.30268].

These formulations differ in what is made explicit: sensed force, tactile contact state, contact phase, desired contact, contact-event ordering, or contact-induced dynamics. This suggests that “explicit” does not denote one canonical variable; it denotes the decision to expose contact structure directly to learning or control rather than leaving it latent.

## 2. Explicit contact observables and sensing pathways

A major branch of contact-explicit policy learning makes contact observable through force, torque, or tactile signals. FACTR 2 addresses the absence of dedicated force sensors on commodity manipulators by learning Neural External Torque Estimation (NEXT) from onboard signals $q$, $\dot{q}$, $q_d$, and motor current $I_m$, with $\tau_m = K I_m$, $\hat{\tau}_f = f_\theta(x)$, and $\hat{\tau}_{ext} = \tau_m - \hat{\tau}_f$ [2606.12406]. NEXT trains in approximately 1 minute from approximately 10 minutes of free-motion data, runs in approximately 1.76 ms per forward pass on a single CPU thread, and on Franka Panda reports contact L1 error $0.547 \pm 0.348$ Nm, compared with $1.471 \pm 0.761$ for a Disturbance Observer and $4.395 \pm 1.531$ for FILIC; in free space it reports $0.414 \pm 0.278$ Nm, compared with $0.449 \pm 0.208$ for the Franka external sensor baseline [2606.12406].

Other works rely on direct tactile or wrist force sensing. NCF-v2 estimates extrinsic contact probabilities over object-surface query points from DIGIT tactile images, recent end-effector poses, and Neural Descriptor Field shape features; in simulation its MSE improves from $0.048$ for NCF-v1 to $0.039$ for the Transformer-based NCF-v2, and the resulting real-world mug-in-cupholder policy reaches $60\%$ success over 15 trials versus $27\%$ for proprioception-only and tactile baselines [2309.16652]. ManipForce records gravity-compensated wrist F/T at $200$ Hz and RGB at $30$ Hz, then uses frequency-aware and modality-aware embeddings in FMT; across six real tasks, FMT reaches an average success rate of approximately $83\%$, compared with approximately $22\%$ for the RGB-only baseline [2509.19047]. MiTaS combines RGB, GelSight Mini, and event-based Evetac, with sensor rates of $25$ Hz, $25$ Hz, and $200$ Hz respectively, and reports average success of $80\%$, compared with $31\%$ for vision-only and $54\%$ for a visual-tactile baseline [2606.06281].

Contact can also be reduced to a stable binary state for dynamics modeling. "Contact-Aware Neural Dynamics" defines real-world hand-level contact by thresholding calibrated fingertip forces from XHand tactile arrays: a fingertip is in contact if $|F_x| + |F_y| + |F_z| > 0.3\,\mathrm{N}$, and the hand-level contact is active if any fingertip exceeds threshold [2601.12796]. The important point is methodological rather than sensor-specific: contact-explicit policy learning depends on making contact information operationally available, but the literature achieves that through sensorless residual estimation, intrinsic tactile perception, wrist F/T, event-based tactile streams, or binary tactile event abstractions.

## 3. Training strategies that elevate contact during policy learning

Once contact is observable, recent methods differ mainly in how they force the learner to use it. FACTR 2 uses external torque not only as policy input but also as a data-selection signal. It defines a scalar contact score $f_t = \|\hat{\tau}_{ext,t}\|_1$, applies hysteresis thresholding to obtain binary contact, and labels each timestep as Free Space, Pre-Contact, or Contact [2606.12406]. FIRST then samples datapoints with probability $p_t = w(s_t)/\sum_j w(s_j)$, typically using $w_F:w_{PC}:w_C = 1:5:1$, or $1:3:3$ for Cap Screwing. Across five long-horizon tasks, FIRST outperforms prior force-aware policies by over $17\%$ in task progress; in the up-sampling ablation, C-only yields $0.670$, PC-only yields $0.818$, and PC+C yields $0.811$, indicating that pre-contact windows are especially influential [2606.12406].

FACTR uses a different mechanism: a force-attending curriculum that corrupts visual input early in training so that a multimodal transformer must rely on force tokens before high-fidelity visual discrimination is restored [2502.17432]. With the same demonstrations and architecture, ACT (Vision-Only) reports average test success of $21.3\%$, ACT (Vision+Force, no curriculum) reports $61.2\%$, and FACTR reports $87.5\%$; the paper summarizes this as a $43\%$ improvement in generalization to unseen objects over baselines without curriculum [2502.17432]. Here, contact-explicitness is not a new observation variable but a training protocol that changes modality saliency.

Reactive architectures schedule contact use in time. PhaForce predicts a contact probability and phase belief with CAP, uses dual-gated visual-force fusion and orthogonal residual injection in a Slow diffusion planner, and routes phase-specific residuals through a Fast corrector at control rate [2603.08342]. Across Charger Plug-in, USB Plug-in, Drawer Opening, and Wiping in both ID and OOD settings, PhaForce reports an average success rate of $86\%$, which is $+40$ percentage points over the average of baselines; ablations removing phase belief, ORI, or Fast all degrade performance substantially [2603.08342]. Tube Diffusion Policy reaches the same issue from a different angle: it learns a feedback flow around nominal diffusion actions, so visual-tactile observations can correct actions step by step within an action tube [2604.23609]. On real-world on-table reorientation and jar opening, TDP reports $96\%/96\%$ success versus $60\%/84\%$ for Diffusion Policy, while using approximately $0.008$ s denoising and $0.003$ s streaming per cycle [2604.23609].

HDP makes contact explicit hierarchically rather than reactively. A high-level Guider predicts objective contacts, and a low-level Actor is trained with both behavioral cloning and short-horizon Q-learning to reach or maintain those contacts [2411.12982]. Across six tasks, HDP reports an average success-rate improvement of $20.8\%$ over Diffusion Policy, and prompt guidance can raise Move-T success from $0.29$ to $0.71$ by replacing the predicted contact with a human-specified one [2411.12982].

## 4. Contact-conditioned dynamics models and model-based reinforcement learning

A second major branch embeds contact explicitly in the dynamics model or the planner rather than only in the policy input. "Contact-Aware Neural Dynamics" uses a simulator prior plus a learned contact-conditioned residual, with a contact predictor feeding a diffusion-based pose model [2601.12796]. On single-object prediction it reports MSE $0.0094$ on real data and $0.0082$ after real fine-tuning, with ADD-S improving from $81.34\%$ to $88.23\%$; on multi-object prediction it reports MSE $0.0075$ and $0.0058$, with ADD-S improving from $73.33\%$ to $79.12\%$ [2601.12796]. For long-horizon prediction, Real-only success is $52.6\%$ on single-object and $47.1\%$ on multi-object settings, whereas Sim+Real w/ Contact reaches $73.7\%$ and $64.7\%$ respectively [2601.12796]. In this formulation, contact-explicit learning primarily serves sim-to-real alignment and policy refinement.

AHAC makes contact explicit through stiffness-aware backpropagation in differentiable simulation [2405.17784]. The method identifies stiff-contact regions using a normalized contact Jacobian and adapts the model-based horizon so gradients are not propagated through unstable contact segments. Across locomotion tasks, AHAC reports about $40\%$ higher asymptotic reward than PPO; on Ant it achieves $41\%$ higher reward than PPO, and on SNU Humanoid it achieves $64\%$ more reward than PPO [2405.17784]. Here, the explicit contact object is not a sensor measurement but a contact-sensitive criterion for deciding where first-order gradients remain reliable.

DAPL instead learns a world model of contact-induced object motion in clutter. Its per-point state includes position, mass, and velocity, and its world model is trained with position, velocity, and velocity-variance objectives so contact-driven dynamic regions are preserved [2603.09882]. On unseen cluttered scenes, success rates are $71.88\%$ in Sparse, $51.04\%$ in Moderate, and $44.56\%$ in Dense settings, and the paper reports over $25\%$ improvement over prior representation-based policies; real-world success reaches $48\%$ across 10 scenes [2603.09882]. ConCent makes the same principle more prescriptive: it extracts a contact event sequence from one real demonstration, optimizes contact geometry so replay explains the observed transitions, and trains RL with stage-gated contact rewards [2606.30268]. On a shape-sorter insertion task with a 40 mm block and a 42 mm square hole, ConCent reaches $80.0\%$ success $(16/20)$, compared with $20.0\%$ without contact geometry optimization, $50.0\%$ without contact-event reward, and $30.0\%$ for unconstrained RL [2606.30268].

Structured contact-explicit dynamics can also appear in simplified analytic models. LCS-RL learns a Linear Complementarity System and then optimizes its parameters through PPO so the learned MPC policy maximizes task performance rather than merely minimizing prediction error [2310.09714]. Across diverse YCB objects, the reported peak success improvements over prior LCS learning range up to $15\%$, and some tasks exceed $70\%$ success using under 30 minutes of data [2310.09714].

## 5. Contact as a task interface rather than a sensor channel

Several recent works treat contact not only as an observation but as the language in which tasks are specified. "Learning to Act Through Contact" defines a goal-conditioned policy over a sequence of desired contacts, with per-end-effector target locations $p^{con}_{t,e}$, binary indicators $I^{con}_{t,e}$, and contact phases Reach, Hold, and Detach [2510.03599]. The same representation is used for quadruped and humanoid locomotion and for humanoid bimanual manipulation. In extrapolated manipulation settings, the contact-explicit policy reports Repose position error $0.115 \pm 0.003$ m versus $0.129 \pm 0.003$ m for a one-hot baseline, Repose rotation error $0.390 \pm 0.006$ rad versus $0.455 \pm 0.007$ rad, and Reorient rotation error $0.109 \pm 0.002$ rad versus $0.191 \pm 0.004$ rad [2510.03599]. The same paper also reports that multi-gait policies achieve the lowest contact plan deviation across durations and gaits [2510.03599].

ContactMimic pushes the same idea into humanoid object interaction by conditioning on explicit part-level binary contact commands [2607.08742]. The policy can therefore be evaluated under identical keypoint trajectories with contact ON or OFF. In simulation, for Wipe whiteboard the commanded-contact policy reports contact bodies $0.65 \pm 0.45$ and impulse $0.44 \pm 0.37$ N·s, while the OFF variant reports $0.25 \pm 0.25$ and $0.04 \pm 0.06$ N·s; compared with BeyondMimic, ContactMimic reports Wipe whiteboard contact bodies $0.65$ versus $0.01$ and impulse $0.44$ versus $0.00$ N·s, despite MPJPE values of $3.6$ cm and $3.9$ cm [2607.08742]. On real hardware, it demonstrates contact controllability across five motions, including $5/5$ ON and $5/5$ OFF for Wipe whiteboard, and $5/5$ ON and $5/5$ OFF for Sit and squat [2607.08742].

A more classical but still contact-explicit perspective appears in "From a Single Demonstration to a General Policy for Contact-Rich Manipulation" [2605.17601]. There, the policy is a hybrid automaton whose states are environmental-constraint primitives and whose transitions are triggered by making-contact events, breaking-contact events, and gripper events. The system reports over $90\%$ success across seven real-world multi-stage tasks, with many tasks at $10/10$ or $9/10$ after zero, one, or two targeted corrections depending on the setting [2605.17601]. In this formulation, explicit contact is not primarily a sensed feature or a diffusion condition; it is the event structure that organizes the policy itself.

## 6. Limitations, misconceptions, and open technical issues

The literature does not support the view that contact-explicit policy learning requires a dedicated force sensor or a single modeling formalism. NEXT infers external torque from onboard motor-side signals without additional sensing hardware [2606.12406]; NCF-v2 makes extrinsic contact explicit from intrinsic tactile observations [2309.16652]; ManipForce and MiTaS rely on high-frequency wrist F/T or multi-resolution tactile sensing [2509.19047] [2606.06281]; other methods instead introduce explicit contact variables inside dynamics models, rewards, or command interfaces [2601.12796] [2607.08742]. This suggests that the common denominator is not a particular sensor stack but the choice to make contact observable, optimizable, or commandable within the learning system.

The open problems are correspondingly heterogeneous. FACTR 2 notes that torque scaling depends on the motor torque constant $K$, that retraining is needed across arms, and that high-speed, large accelerations, saturation, and complex friction transients can reduce estimator fidelity [2606.12406]. "Contact-Aware Neural Dynamics" states that a single binary $c_t$ cannot encode contact area, slip direction, or distributed forces [2601.12796]. PhaForce identifies task-specific phase definitions, diagonal corrective masks, and the absence of formal stability or passivity guarantees as limitations [2603.08342]. ManipForce still reports lower success on Battery Disassembly and Battery Insertion, at $0.65$ and $0.60$, indicating that dexterous disassembly and longer-horizon insertion remain difficult even with high-frequency F/T [2509.19047]. MiTaS reports that co-training can decrease Key performance by $-35$ when Evetac dynamics disturb GelSight’s in-hand localization features [2606.06281]. ContactMimic currently trains one policy per motion and uses binary part-level labels rather than richer contact commands [2607.08742]. ConCent depends on replay fidelity and on primitive approximations that may miss small geometric features critical to contact [2606.30268].

A plausible implication is that the next stage of the field will not be a single dominant architecture, but tighter integration across explicit contact sensing, phase or event structure, contact-conditioned world models, and reactive control. The papers already point in that direction through richer tactile features, multi-contact extensions, hierarchical contact planners, learned subspace routing, and more structured sim-to-real alignment [2601.12796] [2603.08342] [2606.06281] [2510.03599].

Source: https://www.emergentmind.com/topics/contact-explicit-policy-learning