0xAbhi/rc-car-commands

RC-Car Command Dataset

A hand-authored dataset for training a tiny language model that maps a natural-language command to a structured list of tool calls (LangChain/LangGraph style), to drive an RC car via a NodeMCU.

• File: data/rc_dataset.csv
• Columns: input (natural language), output (JSON tool-call list)
• Rows: 700
• Task type: sequence-to-sequence (translation)

Output grammar

Each output is a JSON array of tool-call objects:

[{"name":"Forward","args":{"duration":2}},{"name":"Turn_Left","args":{}},{"name":"Stop","args":{}}]

Tools

Tool	args	Meaning
Forward	{"duration": n}	drive forward n seconds (n = 1–10)
Backward	{"duration": n}	drive backward n seconds (n = 1–10)
Turn_Right	{}	turn right 90°
Turn_Left	{}	turn left 90°
Stop	{}	stop the motors

Labeling rules (applied to every row)

1. Default duration = 2. If a Forward/Backward command has no explicit time, duration is 2.
2. Word-numbers → digits ("five seconds" → 5). Vague durations ("a bit", "briefly", "a moment") → default 2.
3. Durations are clamped to 1–10.
4. Turns take no arguments.
5. Exactly one trailing Stop is always present — appended even when the user didn't say "stop", and never duplicated when they did.

Composite shape templates (stylized, 90° turns only)

Notation: F(n)=Forward, B(n)=Backward, R=Turn_Right, L=Turn_Left, S=Stop.

Shape	Decomposition
U-turn	F(2) R R F(2) S
Square	F(2) R F(2) R F(2) R F(2) R S
Triangle	F(2) R F(2) R F(2) R S (stylized)
Circle	F(1) R F(1) R F(1) R F(1) R S (short segments)
Star	F(2) R R F(2) R R F(2) R R F(2) R R F(2) R R S
Zigzag	F(2) R F(2) L F(2) R S
Figure-8	F(1) R F(1) R F(1) R F(1) R F(1) L F(1) L F(1) L F(1) L S

Because turns are fixed at 90°, only Square is geometrically exact. Triangle, Circle, Star, and Figure-8 are stylized approximations built from 90° steps. They are easy to retune — just change the template above and regenerate those rows.

Left-handed (counterclockwise) shape variants

The first 500 rows draw all shapes clockwise (Turn_Right). Rows 501–700 add left-handed variants for balance — same templates with Turn_Right↔Turn_Left swapped, triggered by phrasings like "to the left", "going left", "counterclockwise". The Zigzag/Figure-8 variants are the mirror image (start with the opposite turn). This is what brings Turn_Left coverage close to parity with Turn_Right.

Coverage (700 rows)

The original 500 rows (see table below) plus 200 rebalancing rows added to fix tool/duration imbalance:

Category	Rows
Forward (durations, defaults, word-numbers, synonyms)	60
Backward	50
Turn left only	30
Turn right only	30
Stop only	20
2-action compounds	90
3–4 action compounds	80
U-turn	25
Square	25
Circle	20
Triangle	20
Star	15
Zigzag	20
Figure-8	15
— rebalancing rows (501–700) —
Backward primitives + compounds (all durations 1–10)	~90
Forward with long durations (3–10)	~40
Turn-left-heavy compounds	~40
Left-handed shape variants (square/circle/triangle/u-turn/star/zigzag/figure-8)	~30

Resulting balance (tool occurrences across all rows)

Tool	Occurrences
Forward	1029
Turn_Right	707
Stop	700
Turn_Left	496
Backward	211

Duration spread: 2 is still the most common (it's the default), 1 is common (circles/figure-8 use 1s segments), and 3–10 each appear 18–44 times.

Notes for training (next step)

• Treat this as seq2seq. Build two small vocabularies (input words; output JSON tokens) plus special tokens <pad>, <bos>, <eos>, <unk>.
• The JSON output is verbose and regular — tokenize it structurally (tool names, "name", "args", "duration", digits, and each { } [ ] : , as tokens) so the vocabulary stays tiny.
• Hold out a small portion (e.g. 10%) as a test set with phrasings not seen in training to measure real generalization.
• The longest outputs are the Star (16 calls) and Figure-8 (17 calls) — use these to choose your max sequence length for padding.