Tactile AI Is Having a Moment, But the Hard Problems Remain Unsolved
A wave of new research on touch-enabled robot learning looks promising on paper, though the gap between benchmark success and real-world deployment is wider than most papers acknowledge.
Image credit: Lottie animation by Centre Robotics (LottieFiles Free, used with credit). · source
Touch is the neglected sense in robotics. For all the progress in computer vision and language models, the field has largely treated tactile feedback as an afterthought, something to bolt on later rather than design around from the start. This week's batch of preprints suggests that might finally be changing, with at least five separate papers tackling various aspects of tactile perception and gentle manipulation. The work is genuinely interesting. It is also, to be precise, nowhere near as solved as the abstracts might suggest.
Let me explain what I mean. The most ambitious of the new papers is Tabero, from a team that includes researchers at institutions I couldn't fully identify from the preprint alone (the code is on GitHub under NathanWu7, for those who want to dig deeper). arXiv hosts the full paper, which introduces both a benchmark and a model architecture for what they call "gentle manipulation," the kind of careful, force-modulated grasping that humans do instinctively but robots find maddeningly difficult. The headline result is striking: their model reduces average grip force by over 70% when given instructions to be gentle, while maintaining high task success rates.
That sounds impressive, and in a narrow sense it is. But it's worth noting that the benchmark itself was constructed by repurposing existing open-source manipulation trajectories, not by collecting new real-world tactile data. This is a reasonable approach given how expensive and time-consuming tactile data collection remains, but it does mean we're essentially testing whether models can learn to be gentle in simulation-derived scenarios. Whether that transfers to actual physical objects with their messy, unpredictable material properties is a separate question the paper doesn't fully answer.
The sim-to-real gap is, actually, the central concern addressed by a second paper on Center-of-Pressure representations for dexterous manipulation. This work, available at , takes a different tack: rather than trying to preserve all the rich information from tactile sensors, the authors propose a physics-grounded representation that captures where forces are concentrated on the fingertip. The intuition is that CoP (as they abbreviate it) is robust enough to survive the inevitable discrepancies between simulated and real sensor readings, while still carrying enough information for complex tasks like peg-in-hole insertion.
Related coverage
More in AI Models
The company just raised its outlook by a staggering amount, and honestly, I'm trying to figure out if this is real momentum or a peak we're about to fall off.
Sarah Williams · 1 hour ago · 5 min
A $65 billion raise that eclipses OpenAI. I've seen big valuations before, but this one's got me scratching my head.
Robert "Bob" Macintosh · 1 hour ago · 3 min
The private equity giants are seeking additional investors for what would be one of the largest AI infrastructure financing deals to date.
James Chen · 2 hours ago · 4 min
The company that once prided itself on vertical integration is outsourcing its AI brain to a competitor. That's not a pivot, it's a concession.
I know I'm being picky here, but the evaluation tasks, peg-in-hole and ball balancing, are classics precisely because they're well-understood. They're good for demonstrating that something works at all, less good for predicting performance on the kind of varied, open-ended manipulation that would actually matter in deployment. The paper claims zero-shot sim-to-real transfer on a multi-fingered hand, which is genuinely notable, though the sample size of test scenarios is small enough that I'd want to see replication before getting too excited.
What connects these papers, and several others from the past week, is a shared recognition that Vision-Language-Action models, the current darling architecture for robot learning, have a tactile blind spot. VLAs are trained primarily on visual and language data, inheriting representations from foundation models that have never felt anything. Teaching them to incorporate touch is not simply a matter of adding another input modality; it requires rethinking how these models represent and reason about physical interaction.
The TacSE3 paper from arXiv illustrates the challenge well. The authors are trying to track object motion inside a gripper using visuotactile sensors, those cameras-behind-gel-pads that have become popular in recent years. The problem is that the images from these sensors are, frankly, boring. Low texture, few distinctive features, nothing for traditional correspondence-matching algorithms to latch onto. Their solution is to convert the tactile images into a three-dimensional force field and estimate motion from that, which is clever but also highlights how far we are from having general-purpose tactile representations. Every paper seems to invent its own way of processing touch data, and none of them have achieved the kind of standardization that ImageNet brought to vision.
This fragmentation matters because it makes progress hard to measure. When one paper reports 70% force reduction and another reports zero-shot sim-to-real transfer and a third reports improved in-gripper tracking, we have no common benchmark to compare them against. Tabero is explicitly trying to address this by proposing a standardized evaluation protocol, but it remains unclear whether the community will adopt it. Robotics has a long history of benchmark proposals that never quite catch on.
Meanwhile, the broader VLA research community is grappling with problems that go beyond tactile sensing. A paper on contrastive representation regularization, RS-CL, argues that VLA models have representations that are insufficiently sensitive to robotic signals like proprioception. The fix they propose is to add a contrastive loss that aligns the model's internal representations with the robot's actual joint states, using the distances between states as soft supervision. The results on RoboCasa-Kitchen push the state of the art to 69.7%, which the authors frame as significant progress. I'd note that 69.7% still means nearly one in three attempts fails, which is... not great for real deployment. But incremental progress is still progress.
More interesting, at least to me, is the AttenA+ paper from arXiv, which makes an argument I find genuinely compelling. The claim is that current VLA training treats all actions as equally important, but this is physically wrong. In actual manipulation, the slow, careful moments, when the gripper is making contact with an object or threading a needle or inserting a peg, matter far more than the fast, coarse motions of moving through free space. AttenA+ proposes reweighting the training loss based on velocity, paying more attention to low-velocity segments where precision matters.
This is the kind of insight that seems obvious in retrospect but apparently nobody had implemented properly before. The paper reports improvements across multiple benchmarks and real-world validation on a Franka arm, which is more than many papers offer. Whether velocity-based attention weighting is the right way to capture "physical criticality" is debatable, there might be slow motions that don't matter and fast motions that do, but the general principle of non-uniform action importance seems sound.
The final paper worth discussing is BORA, which tackles a problem that keeps me up at night: how do you actually deploy these VLA models on real dexterous hands without breaking things? The answer, apparently, is careful offline-to-online reinforcement learning with human-in-the-loop corrections. The framework first trains a critic offline, then does online adaptation with a human ready to intervene when things go wrong. The results show a 33% improvement in success rate over pure imitation learning, and up to 43% improvement on unseen objects.
These numbers are encouraging, but I want to flag something the paper mentions only briefly: the human-in-the-loop component. This is essentially admitting that fully autonomous dexterous manipulation isn't reliable enough yet. Someone has to be watching, ready to step in. That's fine for research, less fine for deployment at scale. The paper is honest about this limitation, which I appreciate, but it does temper the excitement somewhat.
So where does this leave us? The research direction is clearly correct. Robots need touch, VLAs need to incorporate it, and the sim-to-real gap needs closing. The papers this week represent genuine progress on all three fronts. But I'd caution against reading the abstracts and concluding that gentle manipulation is solved. The benchmarks are narrow, the real-world evaluations are limited, and the fundamental problem of tactile representation remains fragmented.
What I'd want to see next is convergence. A shared benchmark that the community actually uses. A tactile representation that works across different sensor types. And, most importantly, longer-horizon evaluations that go beyond single-task success rates to measure robustness over repeated deployments. The papers this week are promising starts. They are not, yet, the finish line.
One more thing, and this is perhaps overly pedantic even by my standards: several of these papers claim state-of-the-art results, but they're measuring against different baselines on different benchmarks. In a field this young and fragmented, "state of the art" is a somewhat meaningless designation. We don't really know what the art is, let alone what state it's in. That uncertainty is, in a way, what makes this moment exciting. The foundations are being laid. We just don't know yet which ones will hold.
