SPLIT sensors and safe HRI controls

Robotics is getting pushed from “don’t touch the human” toward “work safely through contact.” That sounds small, but it changes almost everything. A robot now ne(arxiv.org)ogic at the same time. Two recent papers hit those two halves from different sides: SPLIT for tactile sensing, posted April 27, 2026, and a March 11, 2026 plannin(arxiv.org)s with conformal risk control for safer human-robot interaction. (arxiv.org) ### What is SPLIT, exactly? SPLIT is not (arxiv.org)simulation method for image-based tactile sensors, especially the DIGIT sensor, that tries to generate realistic touch images without relearning every sensor from scratch. The core trick is to separate two things that usually get mixed together: the actual contact geometry and the sensor’s own optical quirks — lighting, background, and unit-to-unit variation. (arxiv.org) ### Why does that matter so much? Touch data is hard to collect in the real world. (arxiv.org) soft materials, camera effects, and tiny manufacturing differences. That makes tactile learning slow and expensive. SPLIT matters because it claims you can adapt across different DIGIT backgrounds and even transfer to a different sensor family, GelSight R1.5, without full retraining. Basically, it is trying to make tactile data less tied to one exact piece of hardware. (arxiv.org) ### What is the other paper do(arxiv.org)It combines control barrier functions — a standard way to enforce safety constraints during motion — with conformal risk control, which is a way to calibrate uncertainty from prediction errors. In plain English, the robot does not just ask “is this motion safe?” It asks “how wrong might my safety estimate be right now, and how much extra margin do I need?” (arxiv.org) ### Why add risk control to barrier functions? Because humans are messy. A person in a shared wo(arxiv.org)arget. Control barrier methods are good at hard constraints, but the hard part is uncertainty in where the person will be next. This framework adjusts safety margins based on the current interaction context, so the robot can stay conservative when uncertainty is high without freezing all the time. (arxiv.org) ### Are these two advances connected? Yes — even if they solve different layers of the sta(arxiv.org)ontact by making tactile simulation more portable and faster. The risk-aware control work helps robots act under uncertainty with formal safety guarantees. One improves what the robot can feel or train on. The other improves what the robot is allowed to do near people. (arxiv.org) ### What’s the practical bottleneck they’re both addressing? Lab demos usually cheat a little. The sensor is carefully calibr(arxiv.org)dictably. Real deployment breaks all three assumptions. Tactile systems suffer when hardware changes across units, and safety controllers suffer when humans move unpredictably. These papers matter because they go after those exact failure modes rather than just squeezing out a prettier benchmark. (arxiv.org) ### So are robots ready for safe contact work now? Not quite. (arxiv.org)at every real tactile stack suddenly generalizes. The control paper shows gains in human-robot navigation scenarios, not a universal solution for all manipulation and contact tasks. But the direction is the right one — better touch models plus uncertainty-aware safety constraints is a much more realistic recipe than either piece alone. (arxiv.org) ### What’s the bottom line? The interesting shift is not one flashy robot demo. It is(arxiv.org)nse contact more robustly and reason about safety more honestly. SPLIT pushes on the sensing bottleneck. Risk-tunable barrier methods push on the control bottleneck. Put together, that is how close-contact human-robot collaboration starts to move from impressive prototype to something you could actually trust on a factory floor. (arxiv.org)