Research Paper · 2025

Knonik Is All You Need for Scaffolding Robot Learning

60 BC-Transformer trainings across 3 storage modes, 10 tasks, and 2 seeds on the LIBERO-Object benchmark (500 demonstrations, 149,014 RGB frames). Policy, optimizer, and evaluation held fixed — only the data format and dataloader change.

Knonik Lossless

64.25% vs 54.35% HDF5

+9.9 pp · 3.51× smaller

Knonik Lossy

54.83% (tied baseline)

+0.5 pp · 18.23× smaller

Stability

HDF5 worst collapse: 90.5 pp

Lossless max: 61.5 pp · Lossy: 46 pp

Download Full Paper (PDF)

Figure 2 — Throughput (samples/s) and per-batch data-fetch latency (log scale)

Throughput (samples / s)

Batch-fetch latency (ms, log scale)

Knonik's parallel prefetch hides decode overhead — compressed data arrives faster than raw HDF5. Median batch-fetch latency drops from 65 ms → 21 ms despite adding a video decode step.

Background

The Data-Infrastructure Problem

Oblivious format

500 demos → 7 GB at 128×128. At 480p+ teleoperation scale: 2.5 TB/day.
HDF5 ignores temporal and spatial redundancy — compressed-video codecs exploit both.
Teams workaround by downsampling or cutting re-training frequency.

Wrong dataloader

PyTorch DataLoader assumes i.i.d. samples. Robotics data is correlated trajectories.
Single-process sequential reads cannot prefetch behind compute.
No global shuffle → correlated batches → seed-sensitive policies.

Silent GPU tax

37% of GPU wall-clock is idle data-starvation with HDF5 + stock loader.
You pay 1.54× the GPU-hours the model actually needs.
nvidia-smi won't show this — only the inter-step gap reveals it.

Section III

Methodology

Dataset

LIBERO-Object: 10 pick-and-place tasks
50 demos per task · ~149,000 RGB frames
128×128 · dual camera · 126–248 timesteps/ep

Training

BC-Transformer policy (unchanged from LIBERO)
Batch 32 · 50 epochs · AdamW
60 total runs: 3 modes × 10 tasks × 2 seeds
Post-hoc eval: 200 rollouts per task

Table 1 — The 10 LIBERO-Object Tasks

ID	Task instruction
T0	Pick up the alphabet soup and place it in the basket
T1	Pick up the cream cheese and place it in the basket
T2	Pick up the salad dressing and place it in the basket
T3	Pick up the bbq sauce and place it in the basket
T4	Pick up the ketchup and place it in the basket
T5	Pick up the tomato sauce and place it in the basket
T6	Pick up the butter and place it in the basket
T7	Pick up the milk and place it in the basket
T8	Pick up the chocolate pudding and place it in the basket
T9	Pick up the orange juice and place it in the basket

Table 2 — Dataset Footprint

Mode	Format	Size	Ratio
Uncompressed (HDF5)	HDF5 raw uint8	6.93 GB	1.00×
Knonik Lossless	Knonik lossless	1.98 GB	3.51×
Knonik Lossy	Knonik lossy (ul2)	0.38 GB	18.23×

Figure 3 — Storage Footprint (GB)

Uncompressed (HDF5)

6.93

Knonik Lossless

1.98

Knonik Lossy

0.38

Section VI

Results

Headline success rate

Figure 4 — Mean Held-out Success · best ckpt · 200 rollouts · 10 tasks × 2 seeds

Uncompressed (HDF5)

54.35

Knonik Lossless

64.25

Knonik Lossy

54.83

Mean success rate (%). HDF5 std ±34.48 pp — highest of three.

Table 3 — Held-out Success Rate (mean ± std)

Mode	Best (%)	Latest (%)	N
Uncompressed (HDF5)	54.35 ± 34.48	49.33 ± 32.64	20
Knonik Lossless	64.25 ± 22.17	53.83 ± 25.84	20
Knonik Lossy	54.83 ± 25.43	51.00 ± 24.79	20

Lossless beats uncompressed by +9.9 pp on bit-identical data — the delta is entirely the pipeline, not the data content.
Lossy matches the baseline at 18.23× smaller storage and outperforms it on several individual tasks.

Per-task breakdown

Table 4 — Per-task Held-out Success (%) · Seed-averaged · Bold = winner per row

Task	Object	Uncompressed	Knonik Lossless	Knonik Lossy
T0	alphabet soup	64.8	52.8	65.5
T1	cream cheese	60.2	58.5	59
T2	salad dressing	92	89.5	80.2
T3	bbq sauce	73.8	64.8	54
T4	ketchup	38.8	29.2	8.2
T5	tomato sauce	83.5	77.8	68.8
T6	butter	16.8	70.8	34.2
T7	milk	29.2	74.8	45.5
T8	chocolate pudding	39.2	66.2	81.2
T9	orange juice	45.2	58.2	51.5
Mean		54.35	64.25	54.83

Seed stability heatmap

HDF5 shows catastrophic per-seed collapses absent from both Knonik conditions — T8 drops 78% → 0.5%, T9 drops 90.5% → 0%.

Figure 5 — Success by Task × Mode × Seed · Green = high · Red = collapse

Task	Seed 0			Seed 47
Task	HDF5	Lossless	Lossy	HDF5	Lossless	Lossy
T0 alphabet_soup	57.5	68.5	75	72	37	56
T1 cream_cheese	79	65.5	68.5	41.5	51.5	49.5
T2 salad_dressing	92	84.5	66	92	94.5	94.5
T3 bbq_sauce	82.5	73	62	65	56.5	46
T4 ketchup	0	11	0	77.5	47.5	16.5
T5 tomato_sauce	78.5	91	76	88.5	64.5	61.5
T6 butter	32.5	75.5	15.5	1	66	53
T7 milk	19	71.5	62	39.5	78	29
T8 chocolate_pu	78	46.5	89	0.5	86	73.5
T9 orange_juice	90.5	89	74.5	0	27.5	28.5

Section VII

Discussion

Cross-seed stability

Table 5 — |Δ| Between Seed 0 and Seed 47 (lower = better)

Mode	Mean \|Δ\|	Max \|Δ\|	Std
Uncompressed (HDF5)	37.7 pp	90.5 pp	32.4 pp
Knonik Lossless	25.2 pp	61.5 pp	17.4 pp
Knonik Lossy	24.6 pp	46.0 pp	11.0 pp

HDF5's worst-case collapse is nearly double any Knonik condition. High mean + high variance is worse in production than lower mean + low variance — the floor matters more once you deploy.
Likely cause: Knonik's parallel prefetch produces more uniformly mixed sample streams, reducing correlated batches that steer training into seed-dependent local minima.

Pipeline efficiency

GPU utilisation rises from 62.9% → 74.8%. The GPU idles 11.9 pp less per step.
Batch-fetch latency: 67.5 ms → 23.8 ms. 50 epochs complete in 48.9 min vs 56.9 min.
Counterintuitive: compression should add decode latency — Knonik's parallel dataloader hides it so effectively that compressed data arrives faster than raw HDF5.

Figure 6 — GPU Utilisation vs. Idle (step/(step+gap))

Uncompressed (HDF5)

62.9% compute

37.1% idle

Knonik Lossless

74.8% compute

25.2% idle

Knonik Lossy

74.2% compute

25.8% idle

Table 6 — Pipeline Efficiency (averaged across all tasks and seeds)

Metric	Uncompressed	Knonik Lossless	Knonik Lossy
Throughput (samples/s)	102.9	119.6	118.2
Total wall time (min)	56.9	48.9	49.5
GPU util (step/cycle, %)	62.9	74.8	74.2
GPU idle (data-wait, %)	37.1	25.2	25.8
Batch fetch mean (ms)	67.53	23.80	24.25
Inter-step gap mean (ms)	108.62	63.30	65.43
GPU energy (Wh)	133.3	122.9	123.8
Est. cost (USD)	2.90	2.50	2.53