Cross-Embodiment Force-Position Interface

• The Cross-Embodiment Force-Position Interface is an embodiment-agnostic control layer that abstracts kinematic differences to preserve task-relevant interaction semantics.
• It employs diverse substrates such as shared latent codes, task-space hybrid actions, and sensorless contact surrogates to bridge force and position cues across varied systems.
• The approach is validated through applications in teleoperation and dexterous grasping, demonstrating improved safety, transparency, and robust transfer despite heterogeneous dynamics.

Searching arXiv for papers on cross-embodiment force/position interfaces and closely related mechanisms. First, searching for the language–force shared-latent work and direct teleoperation/interface papers. A cross-embodiment force-position interface is an embodiment-agnostic control or representation layer that allows task-relevant motion and interaction information to be transferred across agents with different kinematics, dynamics, sensing, and actuation. In the recent literature, this interface appears in several forms: a shared latent space aligning force trajectories with language, task-space hybrid commands combining motion and wrench intent, tactile shear-field policies that turn human-applied forces into position updates on a position-only robot, sensorless teleoperation schemes that infer virtual force from tracking deviation, calibrated torque-to-fingertip-force pipelines for dexterous hands, and geometry-centered interfaces based on particles or contact-point normals rather than native joint coordinates (Tejwani et al., 4 Feb 2025, Li et al., 16 Mar 2026, Bogert et al., 2024, Atar et al., 14 Jun 2026, Wu et al., 14 Jan 2026). Across these formulations, the common objective is not merely trajectory retargeting, but preservation of interaction semantics: what direction to move, how strongly to load contact, and how to realize that intent safely on a different embodiment.

1. Definition and conceptual scope

Cross-embodiment learning arises because action spaces, workspaces, sensing channels, and compliance properties differ substantially across systems. The cited work spans human arms coupled to a 7-DoF haptic exoskeleton, a 6-DoF heavy-duty hydraulic manipulator, torque-controlled arms, position-only gantries, multi-finger dexterous hands with different coupling structures, and robot hands or human hands represented as particles (Hejrati et al., 20 May 2025, Bogert et al., 2024, Atar et al., 14 Jun 2026, He et al., 3 Nov 2025). In this setting, a force-position interface is useful only if it abstracts away embodiment-specific coordinates while preserving control relevance.

Two distinctions organize the field. The first is explicit versus implicit contact representation. Explicit interfaces expose task-space wrench, joint torque, fingertip force, or calibrated load descriptors; implicit interfaces encode contact through tactile shear, end-effector tracking deviation, particle motion, or contact geometry (Li et al., 16 Mar 2026, Atar et al., 14 Jun 2026, Yan et al., 25 Nov 2025, He et al., 3 Nov 2025). The second is semantic versus purely physical abstraction. Some interfaces are grounded in language or task prompts, such as 16-dimensional shared latent codes for force and phrases, or force prompts injected into a vision-LLM; others remain entirely within geometry and dynamics, such as particle displacements or functional point clouds with normals (Tejwani et al., 4 Feb 2025, Li et al., 16 Mar 2026, Wu et al., 14 Jan 2026).

A recurring theme is that “cross-embodiment” does not require identical state or action spaces. In one line of work, different embodiments are unified by a shared latent code; in another, by task-space hybrid action variables; in another, by calibrated physical units such as N·m or fingertip force; and in another, by geometry-centered representations such as particles or point-normal sets (Tejwani et al., 4 Feb 2025, Atar et al., 14 Jun 2026, He et al., 3 Nov 2025, Wu et al., 14 Jan 2026). This suggests that a force-position interface is best understood as a conserved intermediate representation of interaction, rather than as a fixed controller architecture.

2. Representational substrates

The literature uses several distinct substrates to instantiate a cross-embodiment interface.

Interface substrate	Shared content	Representative formulation
Shared latent code	Force trajectory and phrase semantics	16-D dual-autoencoder latent (Tejwani et al., 4 Feb 2025)
Task-space hybrid action	Motion target plus force/wrench intent	$\tau = J^\top(q)\, f_{ee} + \tau_0$; force-aware MoE (Li et al., 16 Mar 2026)
Sensorless contact surrogate	Virtual force or tactile response	End-effector deviation or tactile shear field (Yan et al., 25 Nov 2025, Bogert et al., 2024)
Calibrated contact channel	Joint torque, fingertip force, load descriptor	$\mathcal{G}_h$, $\mathbf{f}_t$, $\boldsymbol{\gamma}_t$ (Atar et al., 14 Jun 2026)
Geometry-centered interface	Contact-relevant position and orientation	Particles or point-normal sets (He et al., 3 Nov 2025, Wu et al., 14 Jan 2026)

The most direct latent formulation is the dual-autoencoder force-LLM. There, processed 3D force trajectories are resampled to 256 time steps, converted to impulse profiles, flattened to 768 dimensions, and mapped by a feedforward encoder to a 16-dimensional latent vector; phrases are mapped to the same latent space either from a 62-D binary representation or a 150-D GloVe representation (Tejwani et al., 4 Feb 2025). Training combines reconstruction, a Hadsell-type contrastive alignment term, and cross-decoding loss,

$$\mathcal{L} = k_r \mathcal{L}_r + k_z \mathcal{L}_c + k_t \mathcal{L}_t,$$

with corresponding force–phrase pairs pulled together and non-pairs pushed apart. The paper explicitly proposes that replacing the force-only encoder with one operating on force-position trajectories would turn this into a cross-embodiment force-position-language interface (Tejwani et al., 4 Feb 2025).

A second substrate is the task-space hybrid action representation. ForceVLA2 formulates contact-rich manipulation as regulation of both task-space motion and end-effector interaction force, with a Jacobian-based execution layer and “active torque-level modulation” rather than passive compliance alone (Li et al., 16 Mar 2026). In the beyond-human-scale teleoperation setting, the same idea appears as coupled velocity references with independent motion and force scaling, where ideal transparency is expressed by

$$G_{\text{ideal}}= \begin{bmatrix} 0 & \kappa_f^{-1}\ -\kappa_p & 0 \end{bmatrix},$$

so that slave velocity scales with $\kappa_p$ and reflected master force scales with $\kappa_f^{-1}$ (Hejrati et al., 20 May 2025).

A third substrate replaces direct force sensing with sensorless contact surrogates. ACE-F interprets follower end-effector positional deviation as a virtual force signal and regulates feedback magnitude by

$$\text{Factor} = \sqrt{\frac{\alpha \cdot \|\Delta \mathbf{ee}\|^2}{1 + \|\mathbf{v}_\text{cartesian}\|^2}},$$

then renders that signal back to the leader through impedance control (Yan et al., 25 Nov 2025). The collaborative manipulation work on tactile transfer uses GelSlim shear fields

$$\tau = \{v_x^L, v_y^L, \text{div}^L, v_x^R, v_y^R, \text{div}^R\}$$

as the common language between an impedance-capable robot and a purely position-controlled gantry, thereby learning a tactile-to-position mapping that emulates compliant behavior without force sensing on the target platform (Bogert et al., 2024).

A fourth substrate makes contact explicit in physical units. In compliant cross-hand grasping, each hand has a retargeting map $\mathcal{G}_h$0 from a shared 15-D MANO-based pose latent and a hand-specific torque predictor $\mathcal{G}_h$1 from raw effort history to joint torque in N·m (Atar et al., 14 Jun 2026). Joint torques are mapped to fingertip forces through the translational Jacobian, and per-finger spatial torque descriptors

$\mathcal{G}_h$2

encode load centroid, participation ratio, and lateral share. This is one of the clearest explicit force-position interfaces in the cited corpus because both motion intent and contact loading are hand-agnostic by construction (Atar et al., 14 Jun 2026).

Finally, several papers define geometry-centered substitutes for direct wrench representations. Cross-embodiment world models represent end-effectors and objects as particle sets and actions as particle displacements, learning a shared dynamics model in particle space (He et al., 3 Nov 2025). CEI represents end-effectors as point-normal sets and aligns trajectories through Directional Chamfer Distance, thereby matching contact-relevant geometry across embodiments (Wu et al., 14 Jan 2026). Neither system exposes force explicitly, but both are designed so that contact effects are preserved through geometric interaction.

3. Architectural realizations and control laws

Architecturally, the cited systems differ in whether the interface is realized as a latent embedding module, a high-level semantic conditioner, or a closed-loop controller.

The force-LLM is the most compact latent-space realization. It has two encoders, two decoders, and a shared latent $\mathcal{G}_h$3 supporting force reconstruction, phrase reconstruction, force-to-language translation, and language-to-force translation (Tejwani et al., 4 Feb 2025). The paper’s interpretation is particularly relevant to cross-embodiment settings: different robots could learn their own encoders into the same latent space and their own decoders back to feasible trajectories, while phrases such as “gently up” or “sharp backward-right” become shared latent points (Tejwani et al., 4 Feb 2025).

ForceVLA2 uses a more hierarchical architecture. A VLM expert processes images, language, and force prompts, while an action expert built as a Cross-Scale Mixture-of-Experts fuses these force-aware task concepts with real-time interaction forces to produce hybrid commands (Li et al., 16 Mar 2026). The policy-level “hybrid” definition is explicit: motion-related and force-related outputs are produced jointly, then realized through a Jacobian-based mapping and torque-level modulation. Force prompts act as semantic switches between non-contact Position Mode and contact-rich Hybrid Mode, so the interface spans semantics, perception, and control in a single stack (Li et al., 16 Mar 2026).

Teleoperation systems implement the interface as a bidirectional coupling law. In the beyond-human-scale hydraulic teleoperation framework, master and slave required end-effector velocities are modified by force feedback,

$\mathcal{G}_h$4

and joint references are obtained through the master and slave Jacobians (Hejrati et al., 20 May 2025). Motion scaling up to $\mathcal{G}_h$5 and force scaling up to $\mathcal{G}_h$6 are integrated into the bilateral coordination law, while sensorless force reflection and a human-robot augmented dynamic model preserve stability and transparency under up to 150 ms one-way delay (Hejrati et al., 20 May 2025).

ACE-F realizes a lighter-weight cross-embodiment teleoperation interface. The leader is treated as a generic 6-DoF end-effector device, with 3-DoF position from a foldable arm and orientation from an IMU or glove; the follower is retargeted by augmented IK, and virtual force is synthesized from end-effector tracking error rather than direct force sensing (Yan et al., 25 Nov 2025). This design separates task-space human intent from embodiment-specific joint-space realization and deliberately limits haptic feedback to 3-D translation, which the paper presents as sufficient for many contact-rich tasks (Yan et al., 25 Nov 2025).

In dexterous grasping, the low-level controller itself is hybrid. A shared pose latent provides a reference configuration, while calibrated torques produce desired fingertip force targets. The force-limited position controller retreats from the position reference when torque overshoot is detected: $\mathcal{G}_h$7 with retreat driven only by positive torque overshoot (Atar et al., 14 Jun 2026). This one-sided compliance is chosen for safety, especially with non-backdrivable actuators, and the same calibrated signal is used during teleoperation, training, and autonomous execution (Atar et al., 14 Jun 2026).

4. Empirical evidence and evaluation

Empirical validation spans translation quality, success rates in contact-rich manipulation, teleoperation transparency, sense of embodiment, and transfer ratios.

For force-language embedding, in-distribution evaluation uses 90% train / 10% test random splits over 30 random seeds, with out-of-distribution evaluations that hold out modifiers or direction words (Tejwani et al., 4 Feb 2025). Dual autoencoders outperform direct MLP baselines across force and phrase metrics, with roughly 20–30% improvement versus baselines. DAE_B achieves the highest phrase-level performance, with ModSim approximately 0.58 and PhraseSim approximately 0.78, while GloVe-based models generalize better to unseen modifiers or directions in force reconstruction and force-direction accuracy (Tejwani et al., 4 Feb 2025). These results matter because a cross-embodiment interface must tolerate embodiment-induced distribution shift, and the paper explicitly interprets out-of-distribution generalization as a necessary property for that role.

ForceVLA2 provides the clearest task-level evidence for hybrid force-position control. Its dataset contains 1,000 trajectories over 5 contact-rich tasks, including wiping, pressing, and assembling (Li et al., 16 Mar 2026). In the comparison table, ACP achieves 16% average success, $\mathcal{G}_h$8 with force input 17%, $\mathcal{G}_h$9 with impedance 22%, and Foca-VLA 66%, with task-specific scores of 80.0 on Press bottle, 75.0 on Clean vase, 70.0 on Clean board, 35.0 on Retrieve plate, and 70.0 on Assemble gears (Li et al., 16 Mar 2026). The paper attributes these gains to active torque-level modulation, force prompts, and force-aware MoE gating rather than to naive force concatenation.

The beyond-human-scale teleoperation system reports high accuracy tracking under up to 1:13 motion scaling and 1:1000 force scaling, and establishes the stability-transparency tradeoff for motion tracking and force reflection under up to 150 ms of one-way fixed and time-varying delays (Hejrati et al., 20 May 2025). In free motion, task time decreases from 35 s at $\mathbf{f}_t$0 to 16.6 s at $\mathbf{f}_t$1, while max position error rises from 0.011 m to 0.029 m; in the user study, the overall normalized sense of embodiment score is 76.4%, with 10 participants and no reported gender limitation (Hejrati et al., 20 May 2025). These data show that a cross-embodiment force-position interface can be evaluated not only by task success but also by transparency and subjective embodiment.

ACE-F emphasizes teleoperation usability and downstream imitation learning. In blind can insertion, ACE-F achieves 100% success versus 50% for Gello, while in simulated box stacking it is reported as approximately 55% faster with fewer errors and higher success rate than joint-copy Gello (Yan et al., 25 Nov 2025). The system is also designed so that demonstrations collected with force feedback improve learned policy performance, including a reported 95% versus 60% success gap in the lifting example (Yan et al., 25 Nov 2025).

Explicit cross-hand force-position interfaces show strong transfer gains. MARC raises cross-hand mean success from 0.44 to 0.71 over 10 reach-and-lift tasks, with improvements such as 0.54 to 0.84 on Inspire and 0.48 to 0.79 on Wuji; removing fingertip force and spatial torque descriptors drops average success from 0.89 to 0.39 in the ablation table (Atar et al., 14 Jun 2026). On an unseen 15-DoF configuration, transfer is achieved zero-shot once a new retargeting map and torque regressor are provided (Atar et al., 14 Jun 2026). This directly supports the claim that calibrated contact feedback, not motion retargeting alone, enables transferable compliant grasping.

Adjacent geometry-centered interfaces also exhibit strong transfer. CEI transfers data and policies from a Franka Panda to 16 embodiments across 3 tasks in simulation and supports bidirectional transfer between UR5+AG95 and UR5+Xhand across 6 real-world tasks, with an average transfer ratio of 82.4% (Wu et al., 14 Jan 2026). The particle-based cross-embodiment world model reports three findings: scaling to more training embodiments improves generalization to unseen ones, co-training on simulated and real data outperforms training on either alone, and the learned models enable effective control on robots with varied degrees of freedom (He et al., 3 Nov 2025). Although these systems are not explicit force interfaces, they supply quantitative evidence that embodiment-invariant intermediate spaces can be learned at scale.

5. Applications and adjacent formulations

The most immediate applications are contact-rich manipulation, collaborative manipulation, teleoperation, and dexterous grasping. ForceVLA2 targets pressing, wiping, retrieving, and assembling under alternating contact and non-contact phases (Li et al., 16 Mar 2026). The beyond-human-scale teleoperation framework targets heavy-duty manipulation with immersive VR, distributed haptics, and force reflection (Hejrati et al., 20 May 2025). ACE-F is positioned as a portable teleoperation platform for collecting high-quality demonstrations across embodiments, especially for contact-rich tasks such as stacking, dragging, wiping, blind insertion, and can sorting (Yan et al., 25 Nov 2025). The tactile collaborative manipulation work addresses cooperative object carrying and box placement, with tactile sensing acting as the common interface between an impedance-capable source robot and a position-only target robot (Bogert et al., 2024).

A second application class is dexterous manipulation under occlusion or deformability. The calibrated force-position interface across heterogeneous hands is evaluated on rigid and compliant objects including toy fruits, a marker, stacked cups, brioche buns, and an egg, and the resulting learned primitives are stated to be reusable in long-horizon manipulation pipelines (Atar et al., 14 Jun 2026). This is significant because the object regimes that benefit most from explicit contact feedback are precisely those where vision is unreliable or motion-only retargeting is underdetermined.

A third application class is semantic mediation between human-readable commands and embodiment-specific execution. The force-language latent model shows that force profiles and phrases can supplement, integrate, and substitute for one another, so that a force trajectory can decode to a phrase and a phrase can decode to a force trajectory (Tejwani et al., 4 Feb 2025). BLM$\mathbf{f}_t$2 is not a force-position interface, but its intent-bridging architecture shows how a frozen multimodal LLM can provide shared high-level semantics to a policy module with embodiment-specific state and action encoders across four robot embodiments and six tasks (Tan et al., 28 Oct 2025). This suggests that semantic intent layers and force-position interfaces are complementary rather than competing abstractions.

Finally, two geometry-centered formulations define an adjacent but important design frontier. CEI uses functional similarity in 3D contact geometry via Directional Chamfer Distance over point-normal sets, while the cross-embodiment world model uses particles and particle displacements as a shared state-action space (Wu et al., 14 Jan 2026, He et al., 3 Nov 2025). Both systems currently encode contact implicitly. A plausible implication is that they provide the structural scaffold on which explicit force or wrench channels could be added without abandoning embodiment invariance.

6. Limitations, misconceptions, and open problems

A common misconception is that a cross-embodiment interface must expose force through direct force/torque sensing. The cited work does not support that claim. ACE-F derives virtual force from end-effector positional deviation, and the tactile transfer system achieves compliant collaborative behavior on a target robot with no force/torque sensing by mapping tactile shear directly to position commands (Yan et al., 25 Nov 2025, Bogert et al., 2024). These are still force-position interfaces in the functional sense, because they preserve contact-aware behavior across embodiments.

A second misconception is that pose retargeting or shared joint representations are sufficient for contact-rich transfer. Several papers argue otherwise. ForceVLA2 explicitly attributes failures of baselines to position-only control or naive force concatenation, while the dexterous grasping work shows that removing calibrated force and load descriptors collapses success toward motion-centric baselines (Li et al., 16 Mar 2026, Atar et al., 14 Jun 2026). CEI likewise notes that visual-kinematic inputs alone limit detection of unstable contacts such as slippage in InsertFlower, and proposes tactile integration as a future direction (Wu et al., 14 Jan 2026).

The current limitations are substantial. The force-language embedding work is restricted to simple single-phase 3-D force interactions and a modifier-plus-direction phrase template, with no position or pose, no torque or wrench, and no temporal models beyond MLP autoencoders (Tejwani et al., 4 Feb 2025). ForceVLA2 avoids heavy sim pretraining because friction and contact modeling are unreliable in simulation, and its dataset, while force-rich, is still modest relative to vision-only corpora (Li et al., 16 Mar 2026). ACE-F renders only 3-DoF translational virtual force, not full 6-DoF wrench feedback, and its virtual force signal depends on controller tuning because it is inferred from tracking error rather than measured directly (Yan et al., 25 Nov 2025).

Calibration cost is another major obstacle. The dexterous grasping interface requires per-hand system identification to train $\mathbf{f}_t$3, and errors in calibration propagate to both policy observations and low-level controllers (Atar et al., 14 Jun 2026). The tactile collaborative manipulation system is currently planar and reports poor normal-force estimation with GelSlim-based setups, limiting straightforward extension to full 3D force control (Bogert et al., 2024). Geometry-centered systems solve embodiment mismatch in position space but omit explicit force, tactile, or impedance channels, which leaves them vulnerable in slip-prone or highly dynamic contact regimes (He et al., 3 Nov 2025, Wu et al., 14 Jan 2026).

Open problems are correspondingly clear in the cited literature. They include learning embodiment-invariant latent spaces with explicit domain adaptation or multi-domain training, integrating richer language and temporal models, using task-centric or object-centric coordinate frames, extending tactile and calibrated force representations to full wrench spaces, incorporating safety constraints directly into controllers and learned decoders, and building multi-robot datasets rich enough to span both kinematic and contact diversity (Tejwani et al., 4 Feb 2025, Li et al., 16 Mar 2026, Atar et al., 14 Jun 2026). The broad technical direction is already visible: position-only abstractions are being supplemented by contact-aware signals, and embodiment-agnostic semantics are moving from geometry and latent pose toward physically calibrated loading descriptors. This suggests that the mature form of a cross-embodiment force-position interface will likely be multimodal, physically normalized, and explicitly hybrid, with motion intent, contact state, and execution constraints represented in a single transferable layer.