π-IMLE

Fast Single-Step Action Generation for Generalist Vision-Language-Action Policies

Kian Hosseinkhani^1*, George Shramko¹, Mehran Aghabozorgi¹, Qinhe Peng², Jianing Qian², Tristan Engst¹, Alireza Moazeni¹, Dinesh Jayaraman², Ke Li¹

¹APEX Lab · School of Computing Science · Simon Fraser University

²University of Pennsylvania

{kian_hosseinkhani, george_shramko, maa143, tme3, sam62, keli}@sfu.ca · {pengqh20, jianingq, dineshj}@seas.upenn.edu

^*Corresponding author

Paper Code

Overview of π-IMLE: the frozen VLM backbone feeds into a single-step cIMLE action head (1 forward pass, 55 Hz) instead of a 10-step flow-matching head (15 Hz), yielding a 3.7x speedup.

The problem & our solution. Current VLAs couple a pretrained VLM with an iterative action head that runs 10 sequential denoising steps per action chunk. π-IMLE replaces this with a single-step cIMLE generator—same VLM backbone, 1 forward pass instead of 10, yielding a 3.67× speedup.

Abstract

Vision-language-action (VLA) policies leverage pretrained vision-language backbones for strong cross-task generalization, but their iterative action heads (e.g. 10-step flow matching in π_0.5) create an inference bottleneck that causes stop-and-go robot movement.

π-IMLE replaces the iterative action head with a single-step conditional generator trained via Implicit Maximum Likelihood Estimation (cIMLE). cIMLE provably preserves multimodal action coverage—avoiding mode collapse—while eliminating multi-step sampling entirely. Applied to π_0.5, π-IMLE increases inference frequency 3.67× (55 Hz vs 15 Hz), enabling up to 11× higher action throughput. On the 40-task LIBERO benchmark it achieves the highest average success rate (98.0%). Real-world Franka Panda experiments show smoother motion and 3.9×–6.6× lower VLA wall-clock per episode, outperforming π_0.5 on every task.

The Speed Story

The core bottleneck in VLAs like π_0.5 is algorithmic, not hardware: the 10-step flow-matching loop imposes a ceiling that system-level optimizations alone cannot break. Porting π_0.5 from JAX to PyTorch+compile or hand-tuned Triton kernels yields only modest gains (20–25 Hz). π-IMLE sidesteps this entirely by generating actions in one forward pass, reaching 55 Hz—a 3.67× increase.

Architecture comparison: π0.5 runs the 300M action expert 10 times sequentially; π-IMLE runs it once, achieving >3.7× speedup.

Inference Frequency (Hz)

VLA forward passes per second (NVIDIA L40S). The first three bars are all π_0.5 under different system-level optimizations — faster runtimes help, but the iterative denoising loop caps the gains. π-IMLE removes that loop entirely, reaching 55 Hz (3.67×).

Action Throughput (actions/sec)

Throughput = inference frequency × execution horizon H. π-IMLE’s single-step generation compounds with a longer horizon, increasing action throughput from 5.5× to 11× over π_0.5.

5.5× speedup (H=8 → H=12)

11× speedup (H=10 → H=30)

VLA-Only Wall Clock per Episode

How do faster inference and higher throughput translate to actual task completion? VLA-WC measures the total VLA compute time per successful episode (excluding robot execution and other non-inference overhead). Lower is better.

Simulation (LIBERO)

Method	VLA-WC (s) ↓	Faster (×) ↑
π-IMLE (Ours)	0.10	10.3×
π_0.5 (Baseline)	1.03	1.0×

Real Robot

Task	Ours (s) ↓	π_0.5 (s) ↓	Faster (×) ↑
Pineapple in bowl	18.56	123.0	6.6×
Swap pineapple & cube	31.1	178.3	5.7×
Pineapple in cabinet	99.4	389.4	3.9×
Pineapple on moving plate	16.7	89.5	5.4×

LIBERO-Plus: Robustness Under Distribution Shift

Robustness under distribution shift. LIBERO-Plus applies systematic test-time perturbations across 4 axes — background, robot initial state, language, and layout — at 5 severity levels (L1–L5), stress-testing each policy as conditions drift further from training.

π-IMLE retains π_0.5's robustness across every axis and severity, while baselines collapse as severity increases. OpenVLA-OFT's L1 regression head folds to the median mode under shift, losing diversity exactly when it's most needed; cIMLE instead enforces mode coverage by construction, so π-IMLE inherits π_0.5's multimodal action distribution while still generating actions in a single forward pass.

LIBERO-Plus robustness curves: accuracy vs severity (L1 to L5) for OpenVLA, OpenVLA-OFT, π0, π0.5, and π-IMLE across four perturbation axes (background, robot init, language, layout). π-IMLE tracks π0.5 closely while baselines drop sharply.

Real-World Videos

We evaluate on a Franka Emika Panda with an NVIDIA A6000, using two camera views (wrist + scene). Both policies are trained on DROID and deployed zero-shot on four tabletop manipulation tasks spanning single-step, multi-step, and dynamic-reactivity categories. π_0.5 runs at 15 Hz (H=8); π-IMLE runs at 55 Hz (H=12), still yielding a shorter wall-clock replan interval.

Motion quality. Across all four tasks, π-IMLE produces noticeably smoother motion. The faster inference minimizes robot idle time, while the longer execution horizon reduces replanning discontinuities. Together these eliminate the visible stop-wait-execute pauses of π_0.5.

Static Task Videos

Task 1 — Pineapple in Cabinet multi-step

Pick up pineapple, place inside cabinet, close the door.

Task 2 — Swap Pineapple & Cube multi-step

Remove pineapple from bowl, then place the red cube inside.

Task 3 — Pineapple in Bowl single-step

Grasp a squishy pineapple and place it into a bowl.

Reactive Task Video

Pineapple on a Moving Plate. The plate is driven by a remote-controlled car, so the target is in continuous motion. Success requires the policy to close the perception–action loop quickly and continuously track the plate while reaching. π-IMLE’s 55 Hz single-step inference acts on fresh observations and lands the pineapple on the moving target. π_0.5 at 15 Hz acts on stale observations and repeatedly reaches toward outdated locations, missing the plate.

π-IMLE (Ours) success

Tracks the moving plate continuously and places the pineapple on target.

π_0.5 (Baseline) failure

Acts on stale observations and reaches toward outdated plate positions.

Benchmarks

We report benchmark results in table form only. First we show real-world robot success outcomes, followed by LIBERO simulation performance.

Real-World Robot Results

Success counts out of 10 episodes on a Franka Emika Panda (zero-shot deployment). π-IMLE outperforms π_0.5 on every task.

Task	π-IMLE (Ours) ↑	π_0.5 ↑
Pineapple in bowl	10/10	9/10
Swap pineapple & cube	9/10	8/10
Pineapple in cabinet	8/10	7/10
Pineapple on moving plate	9/10	6/10

LIBERO Simulation Benchmark

LIBERO comprises 40 tabletop manipulation tasks across four suites (Spatial, Object, Goal, Long) evaluated over 50 episodes each. π-IMLE achieves the highest average success rate (98.0%) among all compared methods while simultaneously leading in inference frequency (3.67× π_0.5). No other method leads in both.

Group	Model	H	Spatial	Object	Goal	Long	Avg ↑	Inf. (×) ↑
Iterative Action Head	π_0.5	10	97.2	99.0	97.8	96.0	97.5	1.0
	π_0.5	30	95.0	98.0	96.2	95.2	96.1	1.0
	π₀	–	96.8	98.8	95.8	85.2	94.2	1.12
Autoregressive VLA	π₀+FAST	–	96.4	96.8	88.6	60.2	85.5	—
Autoregressive VLA	OpenVLA-OFT	–	95.2	94.2	95.2	93.2	94.5	0.69
Reducing Model Size	CogVLA	–	99.0	99.0	97.0	95.0	97.0	0.8
	LightVLA	–	98.0	98.0	98.0	95.0	97.0	1.2
	SmolVLA	–	90.0	96.0	92.0	71.0	87.0	1.0
Layer Distillation	Shallow-π_0.5-L9	–	99.0	98.0	97.0	93.0	97.0	1.7
Layer Distillation	Shallow-π_0.5-L6	–	98.0	96.0	94.0	90.0	95.0	2.3
Single-Step (ours)	π-IMLE	10	98.0	99.8	98.2	96.0	98.0	3.67
Single-Step (ours)	π-IMLE	30	97.2	99.0	96.2	95.8	97.1	3.67

Success rates (%) on four LIBERO suites, 50 episodes per task. H = execution horizon. Inf. (×) = frequency ratio vs π_0.5. Best in bold, second-best underlined.

BibTeX

@article{piimle2026,
  title={\pi-IMLE: Fast Single-Step Action Generation for Generalist Vision-Language-Action Policies},
  author={Anonymous Authors},
  journal={arXiv preprint arXiv:xxxx.xxxxx},
  year={2026}
}