Touch begins where vision ends.
Current vision-language-action (VLA) and visuomotor foundation models exhibit a critical vulnerability in contact-rich manipulation: sensory blindness at the moment of contact. When a robot gripper or human hand engages an object, the camera sightline is inevitably occluded by the end-effector itself. Cameras cannot measure normal force distributions, incipient micro-slips, or frictional shear stresses along the contact manifold.
As a result, vision-only policies cannot distinguish between a firm grasp and an impending drop, nor can they regulate compliance when inserting tight-tolerance pegs or manipulating fragile, deformable materials.
The AGON Tactile Glove Suite resolves this fundamental bottleneck by integrating high-density piezoresistive pressure-sensing arrays directly with continuous 25-DoF electromagnetic field (EMF) hand-pose tracking. By capturing dynamic contact isobars synchronously alongside metric 3D kinematics, we provide the ground-truth tactile trajectories needed to train force-aware diffusion policies and reactive physical foundation models.
High-density contact mapping with zero optical occlusion.
Traditional optical motion capture and camera-based marker tracking fail when fingers wrap around handles or interlock during dexterous manipulation. Our electromagnetic field (EMF) tracking geometry measures continuous 6-DoF position and orientation across every joint segment in real time, impervious to visual line-of-sight blockages.
Concurrently, the tactile array logs normal force gradients at 60Hz to 200Hz across all distal phalanges, proximal pads, and palmar contact zones, generating calibrated pressure tensors ready for immediate tensor ingestion.
Three specialized data-collection editions.
Different physical AI collection workflows require different mechanical trade-offs. From sub-millimeter industrial insertion to all-day wearable household capture, each edition is tailored for high repeatability and seamless dataset streaming.
RigidCore Tactile + EMF
Engineered for high-precision robotic assembly, sub-millimeter peg-in-hole insertion, and micro-manipulation tasks. The RigidCore architecture utilizes a reinforced backing that eliminates sensor deformational hysteresis under heavy normal loads, ensuring linear force transduction across the entire dynamic range.
With 842 pressure-sensing points arranged at an ultra-dense 2 mm spatial resolution, the RigidCore model captures intricate contact geometry, edge orientations, and fine surface texture gradients. Combined with 25-DoF EMF tracking, it outputs fully resolved 3D contact wrench distributions directly to your training pipeline.
- →Electronics PCB component insertion
- →Industrial fastener & threading telemetry
- →Surgical instrument manipulation
- →Benchmarking robot tactile hand parity
How tactile data transforms robotic imitation learning.
Integrating tactile observations into diffusion and transformer policies fundamentally changes how autonomous agents interact with the physical world.
Resolving Contact Occlusion
Vision models degrade rapidly when the robot hand obscures the contact interface. Tactile arrays provide direct, un-occluded sensory ground truth right at the shear boundary, allowing the policy to confirm grasp closure before initiating lift sequences.
Incipient Micro-Slip Compensation
High-frequency 60Hz tactile tensors detect localized shear vibrations and center-of-pressure shifts milliseconds before macro-slip occurs. This enables fast-reactive policies (RDP / FARM) to tighten grasp forces adaptively without crushing deformable items.
Sim-to-Real Contact Parity
Physics simulators struggle to accurately model complex friction coefficients and non-linear material deformations. Calibrated real-world tactile datasets ground synthetic simulation policies with empirical normal-shear dynamics and true material compliance.
Zero conversion friction. Stream directly into LeRobot, MCAP, & HDF5.
Every episode collected with the AGON Glove Suite is packaged with hardware-level timestamp synchronization (<2ms drift). Tactile contact matrices are formatted as standardized NumPy/PyTorch tensors, indexed synchronously with 6-DoF end-effector trajectories, joint velocities, and stereo RGB-D streams.
{
"episode_id": "agon_tactile_demo_0942",
"timestamp_sync_drift_ms": 1.4,
"tactile_tensor": {
"points": 842,
"spatial_resolution_mm": 2.0,
"shape": [T, 842],
"dtype": "float32",
"range_newtons": [0.05, 25.0]
},
"kinematics_emf": {
"hand_pose_dof": 25,
"joint_angles_rad": [T, 25],
"wrist_6dof_transform": [T, 4, 4]
},
"egocentric_rgbd": {
"resolution": [1920, 1080],
"metric_depth_mm": [T, 1080, 1920]
}
}Equip your robotics lab with the AGON Tactile Suite.
Order evaluation units for your engineering team, or commission custom data collection programs with our certified operator fleet.