Contact-Explicit Policy Learning

• Contact-Explicit Policy Learning is a paradigm that directly incorporates measurable contact signals—like external torques, tactile states, and contact events—as key variables in robotic policy training.
• It leverages diverse representations (e.g., binary contact commands, phase labels, and event sequences) to condition policies, reshape training data, and optimize reward structures.
• Empirical studies show that using explicit contact cues improves sim-to-real alignment and overall task success across various manipulation and locomotion benchmarks.

Searching arXiv for 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 (Oh et al., 10 Jun 2026, Jing et al., 19 Jan 2026, Wang et al., 2024, Omar et al., 4 Oct 2025, Li et al., 9 Jul 2026, Kim et al., 29 Jun 2026).

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 (Oh et al., 10 Jun 2026)	$\hat{\tau}_{ext}$ and $s_t \in \{F, PC, C\}$	policy input and phase-aware re-sampling
HDP (Wang et al., 2024)	objective contact $C_t$	high-level contact guidance for low-level diffusion actions
Unified contact-goal RL (Omar et al., 4 Oct 2025)	$p^{con}_{t,e}$, $I^{con}_{t,e}$, $S$, $s$	goal-conditioned multi-task locomotion/manipulation
ContactMimic (Li et al., 9 Jul 2026)	per-part binary contact commands $c_t$	runtime contact controllability
ConCent (Kim et al., 29 Jun 2026)	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 (Oh et al., 10 Jun 2026). In "Contact-Aware Neural Dynamics" (Jing et al., 19 Jan 2026), contact is a binary or probabilistic signal $s_t \in \{F, PC, C\}$0, predicted into the future and fused into a contact-conditioned residual dynamics model. In PhaForce, the explicit schedule consists of a continuous contact probability $s_t \in \{F, PC, C\}$1 and a soft phase belief vector $s_t \in \{F, PC, C\}$2, which jointly determine when force enters planning and where corrective residuals are routed (Wang et al., 9 Mar 2026).

Other formulations elevate contact from a sensed variable to a task specification. The multi-task framework in "Learning to Act Through Contact" (Omar et al., 4 Oct 2025) defines tasks through sequences of desired contact positions, timings, and active end-effectors. HDP predicts objective contacts $s_t \in \{F, PC, C\}$3 as per-finger 3D positions with binary contact flags and uses them to guide a low-level diffusion actor (Wang et al., 2024). ContactMimic conditions a humanoid policy on per-part binary contact commands, so the same geometry can be executed with contact either produced or suppressed (Li et al., 9 Jul 2026). ConCent treats the demonstrated contact event sequence as the central real-to-sim-to-real learning signal (Kim et al., 29 Jun 2026).

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 $s_t \in \{F, PC, C\}$4, $s_t \in \{F, PC, C\}$5, $s_t \in \{F, PC, C\}$6, and motor current $s_t \in \{F, PC, C\}$7, with $s_t \in \{F, PC, C\}$8, $s_t \in \{F, PC, C\}$9, and $C_t$0 (Oh et al., 10 Jun 2026). 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 $C_t$1 Nm, compared with $C_t$2 for a Disturbance Observer and $C_t$3 for FILIC; in free space it reports $C_t$4 Nm, compared with $C_t$5 for the Franka external sensor baseline (Oh et al., 10 Jun 2026).

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 $C_t$6 for NCF-v1 to $C_t$7 for the Transformer-based NCF-v2, and the resulting real-world mug-in-cupholder policy reaches $C_t$8 success over 15 trials versus $C_t$9 for proprioception-only and tactile baselines (Higuera et al., 2023). ManipForce records gravity-compensated wrist F/T at $p^{con}_{t,e}$0 Hz and RGB at $p^{con}_{t,e}$1 Hz, then uses frequency-aware and modality-aware embeddings in FMT; across six real tasks, FMT reaches an average success rate of approximately $p^{con}_{t,e}$2, compared with approximately $p^{con}_{t,e}$3 for the RGB-only baseline (Lee et al., 23 Sep 2025). MiTaS combines RGB, GelSight Mini, and event-based Evetac, with sensor rates of $p^{con}_{t,e}$4 Hz, $p^{con}_{t,e}$5 Hz, and $p^{con}_{t,e}$6 Hz respectively, and reports average success of $p^{con}_{t,e}$7, compared with $p^{con}_{t,e}$8 for vision-only and $p^{con}_{t,e}$9 for a visual-tactile baseline (Krohn et al., 4 Jun 2026).

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 $I^{con}_{t,e}$0, and the hand-level contact is active if any fingertip exceeds threshold (Jing et al., 19 Jan 2026). 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 $I^{con}_{t,e}$1, applies hysteresis thresholding to obtain binary contact, and labels each timestep as Free Space, Pre-Contact, or Contact (Oh et al., 10 Jun 2026). FIRST then samples datapoints with probability $I^{con}_{t,e}$2, typically using $I^{con}_{t,e}$3, or $I^{con}_{t,e}$4 for Cap Screwing. Across five long-horizon tasks, FIRST outperforms prior force-aware policies by over $I^{con}_{t,e}$5 in task progress; in the up-sampling ablation, C-only yields $I^{con}_{t,e}$6, PC-only yields $I^{con}_{t,e}$7, and PC+C yields $I^{con}_{t,e}$8, indicating that pre-contact windows are especially influential (Oh et al., 10 Jun 2026).

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 (Liu et al., 24 Feb 2025). With the same demonstrations and architecture, ACT (Vision-Only) reports average test success of $I^{con}_{t,e}$9, ACT (Vision+Force, no curriculum) reports $S$0, and FACTR reports $S$1; the paper summarizes this as a $S$2 improvement in generalization to unseen objects over baselines without curriculum (Liu et al., 24 Feb 2025). 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 (Wang et al., 9 Mar 2026). Across Charger Plug-in, USB Plug-in, Drawer Opening, and Wiping in both ID and OOD settings, PhaForce reports an average success rate of $S$3, which is $S$4 percentage points over the average of baselines; ablations removing phase belief, ORI, or Fast all degrade performance substantially (Wang et al., 9 Mar 2026). 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 (Xue et al., 26 Apr 2026). On real-world on-table reorientation and jar opening, TDP reports $S$5 success versus $S$6 for Diffusion Policy, while using approximately $S$7 s denoising and $S$8 s streaming per cycle (Xue et al., 26 Apr 2026).

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 (Wang et al., 2024). Across six tasks, HDP reports an average success-rate improvement of $S$9 over Diffusion Policy, and prompt guidance can raise Move-T success from $s$0 to $s$1 by replacing the predicted contact with a human-specified one (Wang et al., 2024).

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 (Jing et al., 19 Jan 2026). On single-object prediction it reports MSE $s$2 on real data and $s$3 after real fine-tuning, with ADD-S improving from $s$4 to $s$5; on multi-object prediction it reports MSE $s$6 and $s$7, with ADD-S improving from $s$8 to $s$9 (Jing et al., 19 Jan 2026). For long-horizon prediction, Real-only success is $c_t$0 on single-object and $c_t$1 on multi-object settings, whereas Sim+Real w/ Contact reaches $c_t$2 and $c_t$3 respectively (Jing et al., 19 Jan 2026). 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 (Georgiev et al., 2024). 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 $c_t$4 higher asymptotic reward than PPO; on Ant it achieves $c_t$5 higher reward than PPO, and on SNU Humanoid it achieves $c_t$6 more reward than PPO (Georgiev et al., 2024). 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 (Zheng et al., 10 Mar 2026). On unseen cluttered scenes, success rates are $c_t$7 in Sparse, $c_t$8 in Moderate, and $c_t$9 in Dense settings, and the paper reports over $E=\{e_s\}$0 improvement over prior representation-based policies; real-world success reaches $E=\{e_s\}$1 across 10 scenes (Zheng et al., 10 Mar 2026). 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 (Kim et al., 29 Jun 2026). On a shape-sorter insertion task with a 40 mm block and a 42 mm square hole, ConCent reaches $E=\{e_s\}$2 success $E=\{e_s\}$3, compared with $E=\{e_s\}$4 without contact geometry optimization, $E=\{e_s\}$5 without contact-event reward, and $E=\{e_s\}$6 for unconstrained RL (Kim et al., 29 Jun 2026).

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 (Bui et al., 2023). Across diverse YCB objects, the reported peak success improvements over prior LCS learning range up to $E=\{e_s\}$7, and some tasks exceed $E=\{e_s\}$8 success using under 30 minutes of data (Bui et al., 2023).

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 $E=\{e_s\}$9, binary indicators $\hat{\tau}_{ext}$0, and contact phases Reach, Hold, and Detach (Omar et al., 4 Oct 2025). 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 $\hat{\tau}_{ext}$1 m versus $\hat{\tau}_{ext}$2 m for a one-hot baseline, Repose rotation error $\hat{\tau}_{ext}$3 rad versus $\hat{\tau}_{ext}$4 rad, and Reorient rotation error $\hat{\tau}_{ext}$5 rad versus $\hat{\tau}_{ext}$6 rad (Omar et al., 4 Oct 2025). The same paper also reports that multi-gait policies achieve the lowest contact plan deviation across durations and gaits (Omar et al., 4 Oct 2025).

ContactMimic pushes the same idea into humanoid object interaction by conditioning on explicit part-level binary contact commands (Li et al., 9 Jul 2026). 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 $\hat{\tau}_{ext}$7 and impulse $\hat{\tau}_{ext}$8 N·s, while the OFF variant reports $\hat{\tau}_{ext}$9 and $s_t \in \{F, PC, C\}$00 N·s; compared with BeyondMimic, ContactMimic reports Wipe whiteboard contact bodies $s_t \in \{F, PC, C\}$01 versus $s_t \in \{F, PC, C\}$02 and impulse $s_t \in \{F, PC, C\}$03 versus $s_t \in \{F, PC, C\}$04 N·s, despite MPJPE values of $s_t \in \{F, PC, C\}$05 cm and $s_t \in \{F, PC, C\}$06 cm (Li et al., 9 Jul 2026). On real hardware, it demonstrates contact controllability across five motions, including $s_t \in \{F, PC, C\}$07 ON and $s_t \in \{F, PC, C\}$08 OFF for Wipe whiteboard, and $s_t \in \{F, PC, C\}$09 ON and $s_t \in \{F, PC, C\}$10 OFF for Sit and squat (Li et al., 9 Jul 2026).

A more classical but still contact-explicit perspective appears in "From a Single Demonstration to a General Policy for Contact-Rich Manipulation" (Li et al., 17 May 2026). 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 $s_t \in \{F, PC, C\}$11 success across seven real-world multi-stage tasks, with many tasks at $s_t \in \{F, PC, C\}$12 or $s_t \in \{F, PC, C\}$13 after zero, one, or two targeted corrections depending on the setting (Li et al., 17 May 2026). 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 (Oh et al., 10 Jun 2026); NCF-v2 makes extrinsic contact explicit from intrinsic tactile observations (Higuera et al., 2023); ManipForce and MiTaS rely on high-frequency wrist F/T or multi-resolution tactile sensing (Lee et al., 23 Sep 2025, Krohn et al., 4 Jun 2026); other methods instead introduce explicit contact variables inside dynamics models, rewards, or command interfaces (Jing et al., 19 Jan 2026, Li et al., 9 Jul 2026). 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 $s_t \in \{F, PC, C\}$14, that retraining is needed across arms, and that high-speed, large accelerations, saturation, and complex friction transients can reduce estimator fidelity (Oh et al., 10 Jun 2026). "Contact-Aware Neural Dynamics" states that a single binary $s_t \in \{F, PC, C\}$15 cannot encode contact area, slip direction, or distributed forces (Jing et al., 19 Jan 2026). PhaForce identifies task-specific phase definitions, diagonal corrective masks, and the absence of formal stability or passivity guarantees as limitations (Wang et al., 9 Mar 2026). ManipForce still reports lower success on Battery Disassembly and Battery Insertion, at $s_t \in \{F, PC, C\}$16 and $s_t \in \{F, PC, C\}$17, indicating that dexterous disassembly and longer-horizon insertion remain difficult even with high-frequency F/T (Lee et al., 23 Sep 2025). MiTaS reports that co-training can decrease Key performance by $s_t \in \{F, PC, C\}$18 when Evetac dynamics disturb GelSight’s in-hand localization features (Krohn et al., 4 Jun 2026). ContactMimic currently trains one policy per motion and uses binary part-level labels rather than richer contact commands (Li et al., 9 Jul 2026). ConCent depends on replay fidelity and on primitive approximations that may miss small geometric features critical to contact (Kim et al., 29 Jun 2026).

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 (Jing et al., 19 Jan 2026, Wang et al., 9 Mar 2026, Krohn et al., 4 Jun 2026, Omar et al., 4 Oct 2025).