An old health check in deep learning
Machine learning engineers share an old health check: a properly designed model should at least converge completely on a small dataset. When you build a new model the first thing you do is make it memorise a single small batch. If it cannot manage that, something is wrong in the architecture or the training objective.
Yet when robotic manipulation is learned purely by imitating demonstrations, whether through a diffusion policy or a Vision-Language-Action model, success rates often fail to reach 100% even on a narrowly defined task with plenty of demonstration data. Across the precision manipulation tasks we have reviewed, the reported figures look like this.
Loss comes down, and the success rate stops somewhere short. The gap lies in the training objective. Imitating a demonstration and succeeding at a task are different objectives. Human demonstrations branch: the same job is done several ways. An imitation model averages those branches and produces a motion belonging to none of them, and a small error at one step compounds into a larger one at the next. The model is not being lazy. It was never optimising for success.
The counterexample: the problem was not the robot
There is an interesting counterexample. HIL-SERL, from UC Berkeley, teaches the same physical robot precision assembly, dynamic manipulation and bimanual coordination, at the level of routing a timing belt and inserting parts. The difference is that it does not imitate demonstrations. It combines reinforcement learning with real-time human correction.
It was never a limit of the hardware.
Optimise directly for task success and correct failures on the spot, and the robot reaches 100% within hours. One detail is worth noting. HIL-SERL still learns a policy per task. What it gained was not generality but mastery.
A question of order
This is where two paths in industrial robot intelligence diverge.
Start broad, then pull precision up afterwards
One large model covers a broad range of tasks. To be fair, frontier labs do not bet on pure imitation either. They layer reinforcement fine-tuning, verifiers and retries, and human correction data to raise the success rate.
Finish one task at 100%, then add tasks
Reach 100% on a defined task first, then widen by adding tasks one at a time. It is not glamorous, but it is the only route a floor accepts. A floor is scored on yield, not on benchmark averages.
What is interesting is where generality-first ends up. The last stretch that lifts a success rate to industrial levels turns out to be task-specific work. You may start from generality, but yield comes from mastery of a single task.
LiOps takes the mastery-first route. Our Large Spatial Model generalises perception: it adapts to a site it has never seen from a handful of samples, and we demonstrated a 100% success rate on an autonomous forklift loading task from ten labelled photographs. Execution we keep stubbornly deterministic, holding cumulative error within 1 mm across the whole chain from the coordinates perception hands over to the robot that carries them out.
Flexible perception, unyielding execution.
Our welding scorecard has no entry for mostly fine.
This is not an argument against VLA
General robot intelligence is coming and we are waiting for it. When VLA matures and understands what to do far better, two things will still be needed on top: a layer that turns the real state of this site into data, and a layer that finishes the work to 100%. Those are the two layers LiOps builds.
An age of generality does not remove the yield layer. It simply means more robots working on top of it.
Reference: HIL-SERL, Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning (Luo, Xu, Wu, Levine, UC Berkeley, 2024, arXiv:2410.21845). Diffusion policy and VLA success-rate ranges are industry-reported figures that vary with task and conditions. The forklift few-shot 100% and the 1 mm welding result are our own.