Dexterous Grasping Is Getting Good. Here's What the Coverage Keeps Missing.

The Harder Problem: Hands That Can Actually Think

KPGrasp is about generating good grasps. The second paper I want to flag is about something slightly different and, I'd argue, the more interesting challenge for anyone thinking about real deployment.

The problem it's addressing is this: we've got these large Vision-Language-Action models now, the VLA systems that have shown genuinely impressive zero-shot generalisation in manipulation tasks. But almost all of them are built around simple parallel grippers. Two fingers, open and close. Try to adapt one of those pre-trained models to a high-DoF dexterous hand and you run into what the researchers call a morphology gap. The model's spatial reasoning breaks down, it forgets what it knew, and you end up with something worse than what you started with. Catastrophic forgetting is the technical term, and it's as unpleasant in practice as it sounds.

The InDex framework from the second paper takes a two-stage approach to get around this. Rather than retraining end-to-end, they repurpose the pre-trained model's parallel gripper output as a kind of intent signal, a scalar that tells the second stage what the hand is trying to do, while a separate diffusion head handles the actual finger joint articulations. The spatial reasoning stays intact; the hand control gets layered on top. It's a bit like, well, it's sort of like keeping your experienced driver and just teaching them to operate a different gearbox rather than starting from scratch with someone new.

Results across multi-stage, contact-rich tasks look solid, and the data efficiency claim is meaningful. You don't want to be collecting thousands of dexterous manipulation demonstrations every time you change end-effectors.

What This Actually Means for Warehouses

Look, here's the thing. Both of these papers are simulation-heavy, and the real-world validation in KPGrasp covers 20 objects, which is a reasonable start but not a warehouse SKU catalogue. The gap between a research demo and a system that runs reliably on a pick line, across shift changes, with worn parts and inconsistent lighting, is still substantial. I've seen too many promising demos not make it past pilot.

But the direction of travel is right. The contact loss tuning problem was genuinely holding things back, and if flow matching gives us a cleaner way to generate grasp priors at scale, that matters. And the morphology gap problem for VLA models is real and underreported; InDex's approach to solving it without destroying the base model's generalisation is the kind of practical thinking that eventually makes it into products.

It's too early to say when any of this lands in production hardware. But I'm less skeptical than I was two years ago, which for me is saying something.

James Chen · 2 days ago · 5 min