save_episode fails for shape=(1,) numeric features due to scalar conversion assumptions

[SevenFo]

Ticket Type

What kind of ticket are you opening?

• 🐛 Bug Report (Something isn't working)

Environment & System Info

• LeRobot: current main (tested from source)
• OS: Linux
• Python: 3.12
• datasets: 4.7.0 / 4.8.4
• numpy: behavior differs across versions (failures seen with newer numpy combinations)
• pyarrow: 23.0.1 in isolated repro env

Description

When a numeric feature is declared as shape=(1,), save_episode() can fail depending on dependency stack.

LeRobot maps shape=(1,) numeric features to datasets.Value(...). In the save path, values can end up in (N,1) form before passing to datasets.Dataset.from_dict. This relies on implicit scalar conversion behavior downstream.

In stricter stacks this fails with:
TypeError: only 0-dimensional arrays can be converted to Python scalars

Context & Reproduction

Minimal reproduction at datasets layer (same schema assumption LeRobot uses for shape=(1,)):

import numpy as np, datasets

ft = datasets.Features({"x": datasets.Value("float32")})
cases = {
	"list_shape1": [np.array([1.], dtype=np.float32), np.array([2.], dtype=np.float32)],
	"stacked_N1": np.stack([np.array([1.], dtype=np.float32), np.array([2.], dtype=np.float32)]),
	"list_scalar": [1.0, 2.0],
}

for k, v in cases.items():
	try:
		ds = datasets.Dataset.from_dict({"x": v}, features=ft, split="train")
		print(k, "PASS", list(ds["x"]))
	except Exception as e:
		print(k, "FAIL", type(e).__name__, e)

Observed:

• list_shape1 / stacked_N1: may fail in stricter stacks
• list_scalar: passes consistently

LeRobot-side impact:

• add_frame accepts np.array([x]) for shape=(1,)
• save_episode may fail due to (N,1) -> scalar conversion assumptions

Relevant logs or stack trace

Representative traceback:

File ".../datasets/features/features.py", line 563, in encode_example
	return float(value)
TypeError: only 0-dimensional arrays can be converted to Python scalars

Checklist

• I have searched existing tickets to ensure this isn't a duplicate.
• I am using the latest version of the main branch.
• I have verified this is not an environment-specific problem.

Additional Info / Workarounds

Proposed fix:

• In LeRobot save path, normalize shape=(1,) numeric columns from (N,1) to (N,) before Dataset.from_dict.
• Keep add_frame input contract unchanged (np.array([x]) still accepted).
• Add regression test to ensure scalarized encoding path remains stable.