Six New Papers Show VLA Models Still Can't Handle Long Tasks Without Help

A separate team takes a different approach entirely. The MPVI paper essentially admits that VLA models can't handle long-horizon mobile manipulation on their own and proposes bolting on classical motion planning. The results are striking: 113% improvement in task progress over a baseline VLA on the BEHAVIOR-1K benchmark. Look, when your fix involves adding an entirely separate planning system with frontier exploration and open-vocabulary object detection, that's not a minor patch. That's an acknowledgment that the end-to-end approach has limits.

The failure recovery problem gets specific attention in the SPR framework, which stands for See, Plan, Rewind. The core insight is that VLA models don't know when they've made a mistake. SPR adds explicit progress monitoring against expected milestones and can "rewind" to recoverable states when things go wrong. The paper claims a 5% improvement over the MolmoAct baseline on LIBERO, which sounds modest until you consider that this is specifically measuring robustness under distribution shift, not just average performance.

From my time building hardware at Fanuc, I learned that the interesting problems aren't usually in the happy path. They're in what happens when a part isn't quite where it should be, or a gripper slips, or sensor noise introduces uncertainty. These papers are all, in different ways, trying to address the gap between demo videos and production reliability.

The ELAN4D paper takes yet another angle, adding what the authors call "4D supervision" by predicting future positions of robot keypoints like joints and end-effectors. The technique uses forward kinematics from proprioceptive states, which means it doesn't require external tracking systems. The auxiliary prediction branch gets discarded during inference, so it adds training cost but not deployment complexity. Results show gains specifically under out-of-distribution conditions including camera shifts and background changes.

Then there's the memory problem. Most VLA models are, as one paper puts it, "stateless." They react to the current observation without remembering what happened before. The Notes-to-Self paper adds a language scratchpad where the model can write down information it needs to remember, like object positions or progress toward subgoals. It's a simple idea that apparently helps significantly on memory-dependent tasks. The fact that this wasn't already standard practice tells you something about how the field has been optimizing for benchmark performance rather than practical capability.

Finally, the Hide-and-Seek framework addresses failure detection, which is arguably more important than preventing failures in the first place. If a robot can recognize when it's failing, it can stop, request help, or try something different. The paper frames this as a supervised learning problem using trajectory-level labels rather than expensive step-by-step annotation.

What's missing from all of this is a unified solution. Each paper improves one aspect of VLA robustness while leaving others unaddressed. The Mixture of Horizons technique doesn't help with failure detection. The MPVI system adds complexity that might not be feasible for all deployments. The scratchpad approach requires changes to training data. We don't know yet whether these techniques compose well together, or whether combining them would introduce new failure modes.

The benchmark numbers are also worth scrutinizing. LIBERO, which appears in multiple papers, is a simulation environment. Real-world results, where they exist, tend to be on simpler tasks with fewer variations. That's an ambitious gap to bridge, and I've seen enough spec sheets to know that simulation performance doesn't always transfer.

Still, the volume of work here suggests the field is taking robustness seriously in a way it wasn't two years ago. The question is whether incremental improvements to the VLA paradigm will be enough, or whether the long-horizon problem requires something more fundamental. It's too early to say. But when six research groups independently decide that the same problem needs solving, that's usually a sign that the problem is real and the current solutions aren't working.

Sarah Williams · 3 hours ago · 6 min