Why Your VLA Will Not Grasp: Failure Modes and a Triage Order

Training finished without errors. The checkpoint saved. You run it, and the arm moves toward the object, closes on nothing, and lifts an empty gripper.

This article contains affiliate links. As an Amazon Associate, we earn from qualifying purchases.

This is the state this article is about. This site does not own the hardware, so the triage order below is our own judgement about sequencing. The material it is built from, however, is sourced — in particular a study that classified and counted failures, which is where we anchor. Checked 2 September 2026.

Start below the model

The instinct is to blame the policy. Resist it for one pass, because the cheap checks are underneath.

The official documentation lists the wiring checks: the power supply, the cable between computer and board, and the three-pin cable between board and servo. Depending on the board, a jumper setting matters too. Unglamorous, and it is where a lot of stalled projects actually are.

Calibration belongs in this tier as well. If the calibration was redone, or a servo horn was reseated after recording, the joint angles in the dataset no longer mean what they meant. Replaying a recorded episode on the hardware separates this cleanly: if the replay itself misses the object, the problem is not the model.

Knowing how these models are put together makes it much easier to judge what they can and cannot do.

Hands-On Large Language Models

USD 37.68 on Amazon.com (as of 2026/09/22)

Amazon

楽天市場

ポチップ

Then check the data, before the policy

If the hardware reproduces the recording, the next suspect is the dataset, and the first question is volume. The developers of the small model recorded that 25 episodes was not enough and performance suffered; the recommendation is at least 50 episodes with 10 per object location. If you are under fifty, suspect the count before anything else.

Epochs deserve a look too. The official guidance is that imitation learning usually converges in five to ten epochs. Judge by how many times the data was traversed, not by an absolute step count — a big step number on a small dataset is not more training, it is more repetition of the same fifty demonstrations.

What the published numbers actually look like

There is a study comparing four policies on a low-cost arm. It is a preprint, submitted June 2026, and should be read with that caveat — but it is specific. The authors ran each policy on each of four tasks for 20 independent hardware trials, 320 evaluation episodes in total.

Before reading anything into the ranking, sit with the absolute values. The best average is a little over half. If your policy fails more often than it succeeds, you are not necessarily doing anything wrong — you may be exactly where the published state of the art is.

The task matters more than the policy

The same study reports per-task numbers, and they are far more informative than the averages.

The same policy lands at 95 per cent on one task and 0 to 10 per cent on another. In the face of that spread, comparing averages is close to meaningless.

And the axis that separates the two is not motion complexity. Moving a pen and sorting by colour involve similar arm trajectories. What differs is whether the destination depends on what was seen. Tasks that require changing where you put something based on a visual property are the hard ones, and they are hard for every policy tested.

The practical consequence: if your task requires that kind of decision, switching policies is unlikely to rescue it. Redesigning the task is the lever.

Once a tool starts running steps on its own, reliability becomes a design problem, and this build-along book walks through the patterns.

AI Agents in Action, Second Edition

USD 47.48 on Amazon.com (as of 2026/09/22)

Amazon

楽天市場

ポチップ

One trial proves nothing

Deciding anything from a single attempt will mislead you. In a regime where success runs thirty to fifty per cent, one success and one failure are both consistent with any hypothesis you might hold.

Running 20 trials per condition, as the study did, is too heavy for an individual. At two minutes a trial, comparing five conditions is 200 minutes. A workable compromise:

• At least 5 trials per condition. Never conclude from one
• Change one variable at a time — a new policy and a new task at once tells you nothing
• Write down what you changed and what fraction succeeded. Memory reconstructs these numbers optimistically

A triage order

Putting it together, this is the sequence we would follow, cheapest and most decisive first.

1. Replay a recorded episode on the hardware. If the replay misses, the fault is below the model — wiring, jumper, or calibration
2. Count the episodes. Under 50, or fewer than 10 per object location, and the dataset is the suspect
3. Check epochs rather than steps. Five to ten passes over the data is the expected range
4. Look at a recorded frame and ask whether a person could decide from it. Occlusion at the grasp moment, shadow, missing depth
5. Ask whether the task requires a visual decision. If it does, the published numbers say 0 to 10 per cent, and task redesign beats policy swapping
6. Only now consider changing the policy, and evaluate it over at least 5 trials

Steps one through four are cheap and rule things out decisively. Step six is expensive and rules out very little. Running them in that order is most of what separates a productive afternoon from a frustrating one.

What to take away

• Published averages run about a third to a half. Failing more than half the time is not evidence that you did something wrong
• The same policy hits 95 per cent on one task and 0 per cent on another. Task design dominates policy choice
• The hard axis is whether the destination depends on what was seen, not how complex the motion is
• 25 episodes was recorded as insufficient. At least 50, with 10 per location
• Replay on hardware first. It separates model faults from wiring and calibration faults in seconds
• Never conclude from a single trial. Five minimum, one variable at a time, written down

For recording the dataset properly in the first place, see the teleoperation article. For the training side, fine-tuning SmolVLA. And when hardware trials become the bottleneck, the sim-to-real article looks at whether simulation is worth it at this scale.

Sources

• arXiv, Benchmarking Vision-Language-Action Models on SO-101: Failure and Recovery Analysis (June 2026, preprint)
• Hugging Face, Imitation Learning on Real-World Robots
• Hugging Face, Compute HW Guide for LeRobot Training
• The Robot Studio, SO-ARM100 repository

ブラウザだけでできる本格的なAI画像生成【ConoHa AI Canvas】

swiftwand

X

AIを使って、毎日の生活をもっと快適にするアイデアや将来像を発信しています。 初心者にもわかりやすく、すぐに取り入れられる実践的な情報をお届けします。 Sharing ideas and visions for a better daily life with AI. Practical tips that anyone can start using right away.