New Research Tackles a Problem Most Robot AI Ignores: What Happens When Instructions Are Wrong

A different flavor of the same issue

A separate paper addresses a related problem: the delay between when a robot observes something and when it acts on that observation. DEFLECT, from researchers working on vision-language-action policies, tackles what happens when your fancy large model takes so long to process an image that the world has moved on by the time the robot executes the resulting action.

The technical approach is interesting. Rather than trying to speed up inference (the obvious solution that hits hardware limits), they train the policy to anticipate the mismatch. A frozen reference model generates what the "correct" action would be from the future state, and the trainable policy learns to favor those execution-time-aligned actions even when working from stale inputs.

The improvement: up to 6.4 percentage points in high-latency success rates. On real robot tasks, they report a 4.6 percentage-point gain at the longest delays tested.

Look, those aren't revolutionary numbers. But they're achieved without architectural changes or additional inference-time computation, which matters for actual deployment.

What this means

Both papers point to the same underlying tension in robotics AI right now. The models are getting impressively capable in controlled conditions, but the gap between demo and deployment keeps biting people.

The agricultural navigation work is particularly relevant as companies push harder into farm automation. I've been skeptical of some of the more ambitious claims in this space, and research like this validates that skepticism. If your navigation agent can't handle a farmer saying "go past the, uh, the big tree" when there's no big tree, you don't have a product. You have a demo.

The latency work matters for anyone deploying large VLA models on robots that need to move continuously. The stop-and-go behavior of synchronous execution is obviously bad, but asynchronous execution creates its own problems. DEFLECT offers one approach, though the real test is whether it holds up across more diverse manipulation tasks than those tested.

Neither paper solves their respective problems completely. Both acknowledge limitations. That's actually refreshing. The field could use more work that says "here's a real problem, here's partial progress, here's what we still don't know."

Mark Kowalski · 18 hours ago · 5 min