IROS 2026: Open-Source Tactile Sensor and Data-Efficiency Keynote Point to Engineering Shift

The conference's main papers and workshops pointed in the same direction: embodied AI is moving from language-driven demonstrations toward physical interaction and engineering. Workshops such as Physical World Models for Scaling Embodied AI, Touch-to-Action, Sensors and Actuators for Dexterous Manipulation, and Rethinking Uncertainty for Modern Robotics Paradigms focused on physics, tactile sensing, and uncertainty. The WorldArena 2.0 challenge added visual-tactile perception, reinforcement-learning utility, and real physical manipulation to one evaluation system. Papers including ULTRA, HALO, LITHE, VOFA, KineFuse, VTAP Gripper, and ViHaTeleop addressed physics-aware control, hierarchical brain-cerebellum architectures, dexterous manipulation, and lower-cost data collection.

FlexiTac is a low-cost, open-source piezoresistive tactile sensor. It uses an FPC and force-sensitive film stack, can be assembled in about three minutes, and can be bent, cut, and wrapped around different surfaces for gloves, fingertips, or robot skin. The team released the bill of materials, step-by-step assembly instructions, and software for reading tactile data, and integrated the stack into the LeRobot robot-learning framework. Li said even an undergraduate with a 3D printer capable of making a low-cost SO-100 arm could add the sensor and begin tactile research.

According to Li, about 18 months after the open-source release, FlexiTac has been used in academia and industry for multimodal policies, deformable world models, quadruped mobile manipulation, and VLA tactile feedback. A multi-finger dexterous manipulation project with Purdue University, the VTAP Gripper, was shortlisted for an IROS best paper award. Analog Devices is developing a commercial version with tactile resolution five times that of a human fingertip, using semiconductor manufacturing. Li also described long-term stability: sensors made by a student in summer 2024 still worked about 18 months later, and a model trained on data from a year earlier continued to function. Batch-to-batch consistency was another stated advantage.

The sensor's simple sensing mechanism also makes simulation easier. Li said it can be modeled as a spring-damper system with three hyperparameters, allowing large-scale randomized training and sim-to-real transfer. The same framework was tested in a dual-arm assembly task and on an NVIDIA humanoid robot. For real-world data collection, the team modified the UMI handheld device, which normally records precise pose but not the tactile feedback felt by the human operator, so that it records synchronized vision and touch. The data can be used for pretraining and then for fine-tuning with teleoperation. Tasks included reorienting a transparent test tube under external disturbance, adjusting a pencil in hand and inserting it into a sharpener, controlling gripper opening while pipetting, and maintaining force during whiteboard wiping.

In the question-and-answer session, Li said piezoresistive material cannot directly measure shear force, but high spatial and temporal resolution can approximate it. A very thin soft layer on the sensor surface can show force-distribution shifts under static and dynamic shear. Li said Analog Devices aims for at least five to seven times the current resolution, and two related papers, including one on slip detection with piezoresistive material, were accepted at CORAL. Li also said FlexiTac works well with the UMI gripper because its fingers are compliant and its internal mechanical spring provides compliance for force-sensitive tasks.

Chung's keynote took a different angle: the central problem is not data volume but information density. Using autonomous gliding for fixed-wing drones as an example, she said exploration costs energy and that after a certain airspeed, net energy gain falls to zero, so collecting more data in that region cannot improve the policy. In shortest-path planning through an unknown cost field, sampling should be guided by whether a measurement would change the estimated path, not by uniform coverage. In distributed multi-robot systems, bandwidth, delay, and packet loss make broadcasting all data impractical. Her team transmits the most decision-relevant information instead, uses Bloom filters with random salting to compress map information while preserving Bayesian fusion across cycles, and applies streaming-style progressive transmission so robots can act on partial data.

Chung also criticized end-to-end policy learning. In contact-rich assembly tasks, she said, large multimodal datasets can still produce mediocre deployment results, including a policy that repeatedly performs an insertion motion even when the box is empty. Her team's policy-repair method selects corrective samples and, under the same demonstration budget, improves generalization more than simply adding data. The message, she said, is that data quality sets the upper bound.

The Pittsburgh meeting also showed growing Chinese participation. Papers and workshops included Dalian University of Technology's MIWA torque-assistance method, a dual-wheel-legged mobile manipulation framework called DeLM developed with Kuwa Technology, and a dual-arm manipulation workshop organized by Peking University's Dong Hao team. Leiphone.com reported that invited participants included engineers from the dexterous-hand company Sharpa and Mu Yao, who recently founded SeeAct AI.