Direct Training versus Fine-Tuning

9.1 What direct end-to-end IL requires

ACT with ALOHA: 50 demonstrations per task (100 for the hardest), roughly 10–20 minutes of collection, yielding 80–90% success on six real bimanual tasks. Diffusion Policy: +46.9% average over prior methods across 12 tasks. robomimic PH: 200 trajectories per task for 84–100% on the easier tasks and 71.3% on Transport with BC-RNN.

So direct training is feasible at the 50–200 demonstration scale — which an EMG protocol could produce in a few sessions. The question is whether it should be, and §3–§5 answer no: at 50 episodes, the Belkhale policy-noise penalty is at its most severe (69.7% vs 48.3%), robomimic shows mixed-quality data underperforming smaller proficient sets, and every EMG artifact in §4 manufactures policy noise.

9.2 What fine-tuning requires

Model	Adaptation budget	Compute	Source
OpenVLA	"small datasets with 10–150 demonstrations"	Full FT: 8×A100 for 5–15 h. LoRA: 10–15 h on one A100 — 8× reduction, 1.4% of parameters, 68.2% vs 69.7% full FT	Kim et al., CoRL 2024 [arXiv:2406.09246] [A]
π₀	Post-training 5 h (simple tasks) to 100+ h (laundry, mobile manipulation)	—	[arXiv:2410.24164] [C]
π₀.₅	~400 h across ~100 homes; scaling curve 3→104 locations; 97.6% of phase-1 data is not mobile-manipulator household data	—	[arXiv:2504.16054] [C]
Gemini Robotics On-Device	"as few as 50 to 100 demonstrations" for new tasks	—	DeepMind blog [D]
On-Device 2	"fewer than 200 examples" for a new embodiment; SO101 6.7% → 53.3%, Dexmate 24.4% → 75.6%	"a few hours of training"	Model card, 30 Jul 2026 [D]
GR00T N1	100 demos/task in sim benchmarks; on real GR-1, 10% of teleop data reaches 42.6% vs Diffusion Policy 46.4% on 100%	—	[arXiv:2503.14734] [B]
SmolVLA (450M)	<30k episodes total pretraining; SO-100 real tasks 78.3% with community pretraining vs 51.7% without	Single GPU; CPU deployment	[arXiv:2506.01844] [C]
TRI LBM	"less than 30% of the data needed for training from scratch"; on one real task 15% sufficed	~1,700 h pretraining	[arXiv:2507.05331] [A]

Note carefully: LoRA changes the compute story by 8× but there is no published evidence that it changes the data requirement. It matches full fine-tuning at the same demonstration count; it does not reduce that count.

9.3 The verdict, and the reason

Fine-tuning, and specifically fine-tuning with action-space extension.

The argument is not primarily about dataset size, though size settles it: no realistic EMG protocol produces anything near the ≥1,000-hour scale that pretraining requires, while 50–200 demonstrations is comfortably within one or two sessions.

The argument is about what the data contributes at the margin. Adding EMG-teleoperated trajectories to a corpus as more samples of the same action space is a losing proposition — the samples are noisier, slower and more inconsistent than what a leader-follower rig produces, and §3.2–§3.3 tell us that mixing them in degrades rather than improves. This is the robomimic MH condition and it costs 30 points on Transport.

Adding them as a new action dimension is a different proposition entirely. ForceBand's $a_t \in \mathbb{R}^{11}$ contains a force channel that no other data source in the corpus provides. Tactile-VLA and ForceVLA2 demonstrate what happens when a π₀-class model gains a force output: 5% → 35% on USB insertion, 18.0% → 66.0% average across five contact tasks. The marginal value of the data is high precisely because it is not substitutable.

This also resolves an apparent contradiction with §5.3's finding that modality mixing helps (+20% for kinesthetic plus VR). Modality heterogeneity helps when the modalities contribute orthogonal quality dimensions. EMG-as-pose contributes a strictly worse version of a dimension already covered — pure downside. EMG-as-force contributes a dimension nothing else covers — pure upside, modulo label noise.

9.4 Curation is not optional

Whatever the integration path, the curation literature makes a strong claim: on mixed-quality human data, removing demonstrations improves policies.

• CUPID (Agia, Sinha et al., CoRL 2025) [arXiv:2506.19121] [A]: influence functions estimating each demonstration's causal effect on closed-loop return. "Training with less than 33% of curated data can yield state-of-the-art diffusion policies on RoboMimic." Hardware: Figure-8 ~60% → ~85%, TuckBox ~40% → ~80%, Bookshelf ~45% → ~70%, with 50–66% of demos removed. Also: intrinsic quality heuristics actively degraded performance on two of three tasks — a caution that smoothness-based filtering is not universally safe.
• DemInf (Hejna et al., RSS 2025) [arXiv:2502.08623] [A]: score trajectories by contribution to $I(S;A) = H(A) - H(A|S)$, which simultaneously rewards diversity and penalises conditional action entropy. 5–10% improvement on RoboMimic at a 50% filter.
• Demo-SCORE (Chen, Lessing, Liu & Finn, RSS 2025) [arXiv:2503.03707] [A]: 15–35 absolute points higher success across five real ALOHA tasks by filtering with a rollout-trained classifier.
• RINSE (2026) [C]: +16% on Transport with one-sixth of the data via spectral arc length.
• Re-Mix (Hejna et al. 2024) [arXiv:2408.14037] [C]: group DRO over dataset mixtures — 38% over uniform weighting, competitive using only 25% of the original data.

For EMG-sourced data specifically, this suggests a mandatory pipeline stage: filter by SPARC or by influence before mixing into any training set, and report what fraction survived. Given §4's artifact catalogue, the surviving fraction is itself a useful measurement of medium quality — and it would be the first quantitative comparison of EMG-collected data against conventionally-collected data on the axis that matters.

---