Tactile feedback is widely recognized as a missing ingredient for human-like robotic dexterity. To train complex manipulation policies, roboticists frequently turn to physics simulators where reinforcement learning (RL) agents can safely experience millions of interactions. However, transferring these learned policies from simulation to physical hardware—a process known as sim-to-real transfer—faces a notorious bottleneck: the contact physics gap.
While simulators can easily compute robot kinematics, they struggle to accurately model the microscopic realities of touch. Real-world tactile sensors are plagued by non-linearities, sensor noise, thermal drift, and material hysteresis. Simulating these continuous force dynamics with high fidelity is computationally exhausting and frequently inaccurate.
To solve this, a powerful and counterintuitive idea has emerged: instead of building more complex simulators or higher-resolution force sensors, we should throw away the continuous force data and abstract touch into simple binary contact states. By reducing tactile feedback to a "touch/no-touch" signal, researchers can perfectly align the observation space of a simulator with physical hardware, enabling seamless, zero-shot sim-to-real transfer.
The Primacy of Contact Geometry
The core reason binary abstraction works is a fundamental insight about manipulation: knowing where and when a contact occurs is often far more critical for control than knowing exactly how hard the sensor is being pressed.
In simulation, determining whether two geometric meshes intersect is a clean, deterministic operation. By thresholding real-world sensor data to match this binary abstraction, the policy becomes invariant to the unmodeled physical discrepancies of the real world.
As noted by Liu et al. (2025), simplifying tactile feedback into binary contact signals focuses the learning algorithm entirely on contact geometry. This geometric focus provides sufficient information for complex dexterous tasks while bypassing the need to measure precise force magnitudes. Even when using highly simplified sensor models, Ding et al. (2021) demonstrated that processing tactile signals into binary values via a heuristic threshold provides substantial benefits for physical interactions like door opening.
This abstraction is equally powerful when applied to vision-based tactile sensors (VBTS), which capture high-resolution images of an elastomeric skin deforming. Rather than trying to simulate the complex optical and material deformations of the gel, Su et al. (2024) thresholded the difference between active and reference images to create a binary contact mask. This representation strips away non-essential visual noise like lighting and texture, allowing a policy trained entirely in simulation to generalize to unknown real-world objects.
Filtering Out Physical Non-Linearities
Continuous force sensors do not just report force; they report the physical artifacts of their own construction. Elastomeric materials suffer from hysteresis (where the sensor's output depends on its past states), and friction coefficients change as surfaces wear down.
Binarization acts as a low-pass filter for these complex physical phenomena. As shown by Wu et al. (2019), abstracting tactile data into binary contact states minimizes the discrepancy between simulated physics and real-world sensor noise, leading to highly robust zero-shot deployment on multi-fingered grasping tasks.
Continuous Force Sensing (High Sim-to-Real Gap)
[Real Force + Hysteresis + Noise] ----> Policy struggles to generalize
Binarized Tactile Abstraction (Low Sim-to-Real Gap)
[Real Force + Hysteresis + Noise] ----> [Threshold] ----> [ 0 / 1 ] ----> Policy transfers seamlessly
This discretization can be extended slightly beyond binary states to capture directional intent without reintroducing continuous noise. For instance, Yin et al. (2024) abstracted 3-axis force data into binary normal signals (contact/no-contact) and ternary shear signals (indicating positive, negative, or zero directional slip). This formulation successfully filtered out material hysteresis while preserving the essential directional information required for a robotic hand to translate diverse objects across its palm.
Compensating with Space and Time
If a robot only receives a 0 or 1 from its fingers, how can it achieve high-level dexterity? The loss of continuous depth information is compensated for in two ways: spatial density and temporal history.
On the spatial front, distributing many low-cost binary sensors across a robot's hand allows the policy to perceive the shape and orientation of an object through the spatial pattern of active contacts. Yin et al. (2023) utilized a hand covered in distributed binary touch sensors to achieve complex in-hand rotation of irregular objects like rubber ducks and tomatoes, completely blind. Similarly, Xue et al. (2023) developed a distributed manipulation table (ArrayBot) using a 16×16 grid of tactile-sensing pillars. By using binarized contact maps as the primary observation modality, the system generalized across unseen object shapes with zero real-world fine-tuning.
On the temporal front, a single binary bit changing over time maps out the dynamics of motion. Miller et al. (2025) demonstrated that sparse binary contact signals are sufficient for superhuman dexterity tasks—such as bouncing a ball rapidly on a paddle—when paired with effective temporal representation learning. The policy infers velocity and acceleration not from force magnitude, but from the precise frequency and timing of the binary state switches.
Architectural and Functional Abstractions
Binarization does not have to happen exclusively at the raw sensor level; it can also be integrated into the control architecture itself.
Instead of feeding raw tactile data directly to a monolithic policy, Veiga et al. (2020) proposed a hierarchical decomposition. They abstracted tactile feedback into the operational states of local, low-level finger stabilizers. The high-level policy interacts only with these functional states rather than noisy sensor data, shielding the global policy from the sim-to-real gap.
When tasks demand more than binary data—such as high-precision bimanual assembly—the trend still leans toward mathematical simplicity. Huang et al. (2025) showed that selecting tactile modalities that are mathematically straightforward to model, such as normal force curves, allows researchers to bypass the complexities of optical simulations while retaining necessary feedback. Conversely, where complex deformations are inevitable, Zhang et al. (2025) demonstrated that abstracting vision-based tactile data into geometric contours and squeezed areas provides a direct pipeline from simulation to reality.
Ultimately, the success of binary tactile abstraction across tasks like blind valve rotation (Lee et al., 2024) and in-hand manipulation proves a core tenet of modern robotics: simplifying the representation of the physical world is often the fastest way to master it.
Key Foundational Work
MAT: Multi-Fingered Adaptive Tactile Grasping via Deep Reinforcement Learning (2019) — contributed 2 concept elements to this direction.
Hierarchical Tactile-Based Control Decomposition of Dexterous In-Hand Manipulation Tasks (2020) — contributed 2 concept elements to this direction.