Two New Papers Show How to Fix Robot Policies Without Starting Over

On four bimanual tasks (two-armed manipulation, which is genuinely hard), FlowPRO beat four baselines. The paper doesn't give exact success rate numbers in the abstract, so I can't tell you by how much.

EVE takes the opposite approach. Instead of fixing the policy through additional training, it asks: what if we just made the robot think harder at test time?

This is directly inspired by what's happened with large language models. The whole "test-time compute scaling" thing, where you let a model generate multiple candidates and then verify which one is best, has been transformative for reasoning tasks. arXiv published EVE as an attempt to bring that same idea to robot control.

The setup: you take a frozen base policy (no additional training) and wrap it with multiple VLM-based verifiers. The robot proposes an action, the verifiers score it and suggest refinements, and an "action incorporator" fuses that feedback back into the final motion.

You might be wondering why you'd need multiple verifiers. The paper studies this explicitly, looking at how different verifiers with distinct capabilities can be combined. It's modular in a way that feels practical, you could swap in better verifiers as VLMs improve without touching the base policy.

What I find compelling about EVE is that it requires zero additional training. You're purely spending more compute at inference time. For applications where you can't easily collect correction data (remote robots, dangerous environments), this could matter.

So which approach is better? Honestly, I don't think that's the right question. They're solving different versions of the problem.

FlowPRO is for when you have access to the robot and a human operator who can provide corrections. It permanently improves the policy. EVE is for when you need better performance right now, without any additional training infrastructure.

I initially thought these were competing methods, but after reading both papers, they seem more complementary. You could imagine using EVE for immediate deployment, then collecting failure cases to feed into FlowPRO for periodic policy updates.

What remains unclear is how these scale. FlowPRO was tested on four tasks. EVE was tested across "diverse simulated and real robotic tasks" but the abstract doesn't specify how many or which ones. Neither paper addresses what happens when you need to handle hundreds of different tasks, which is where the real deployment challenges live.

The bigger picture here is that the field seems to be converging on a realization: the hard part isn't initial training anymore. It's everything that comes after. Both papers are essentially post-training infrastructure, ways to take a model that mostly works and make it actually reliable.

This feels like a maturation of the VLA approach. We've moved past "can we train these at all" to "how do we make them production-ready." That's progress, even if neither paper fully solves the problem.

I'll be watching to see if either of these approaches gets adopted by the humanoid companies. They're all sitting on VLA-based systems that work in demos but, as far as I can tell, struggle with the messiness of real deployment. FlowPRO and EVE offer two different paths forward. It's too early to say which one wins out, or if the answer is both.

Robert "Bob" Macintosh · 14 hours ago · 4 min