ENYOLAB · sample-efficient imitation · AMD ROCm / Radeon AI PRO R9700
Under a small budget of demonstrations, does aggressive regularization buy sample-efficiency?
A Diffusion Policy learns the PushT manipulation task from a handful of teleoperated demos. We vary the number of demos and the recipe (standard vs. heavily regularized) and measure how far each gets. Inspired by Konwoo et al. on data-constrained pre-training.
The policy actually solves the task
Every evaluation records the rollout. These are real episodes: the pusher (blue) drives the grey block onto the green target. Episodes end the instant coverage crosses 95%.
standard recipe, 200 demos — diffusion policy on PushT
enhanced recipe, 200 demos — + weight-decay + augmentation
Preliminary evidence: under a 100-demo budget, enhanced regularization improves PushT success from 8% to 19% (+11 pts), while showing no gain at 200 demos.
Data-efficiency signal
Success rate on 100 evaluation episodes, seed 0. The question is whether the enhanced curve sits above standard — i.e. reaches the same success with fewer demos.
- Enhanced (+ weight-decay + augmentation)
- Standard (LeRobot defaults)
Method — the honest spec
Preliminary: single seed, one GPU, a baseline that plateaus below the published reference. Enough to show the signal, not yet to publish.
- Policy
- Diffusion Policy
- Task
- PushT (sim)
- Steps / run
- 150,000
- Eval episodes
- 100
- Seed
- 0
- Throughput
- ~11 step/s
- Wall-clock / run
- ~3h48
- Hardware
- R9700 (gfx1201)
What we haven't done yet — and should
- credibilityMultiple seeds & error barsA standard-vs-enhanced delta on one seed can be noise. 3+ seeds turn the signal into a claim.
- baselineClose the 29% → 65% gapThe one remaining config gap is batch size (8 vs ~64). Gradient accumulation could recover it without a bigger GPU.
- ablationIsolate the leverEnhanced bundles weight-decay + augmentation. Which one drives the gain? Two ablation runs answer it.
- edgeShip it to the JetsonThe north star: export a >29% policy and run few-shot manipulation on the edge — the 30-second demo you can hold.