Muninn: Your Trajectory Diffusion Model But Faster

Authors: Gokul Puthumanaillam, Hao Jiang, Ruben Hernandez, Jose Fuentes, Paulo Padrao, Leonardo Bobadilla, Melkior Ornik · RSS 2026

Licences

Open Access
Paper: CC BY 4.0

Everything below draws on the full paper under its CC BY 4.0 licence; its figure is credited to the authors.

TL;DR

Muninn decides at each denoising step whether to reuse a cached denoiser output, giving up to 4.6x faster diffusion planning with a calibrated bound on deviation from the full-compute plan.

Why it matters

It speeds up existing pretrained diffusion planners and policies without retraining, extra data or architecture changes, and it can be stacked on distilled or smaller models for further gains. The remaining budget is also a runtime signal a controller can use to fall back to full compute or a safer controller.

Problem

Diffusion trajectory planners and diffusion policies run tens of denoising steps per plan, which is too slow for embedded platforms and high-rate control loops. Existing speed-ups either change the sampler or compress the network, which costs plan quality or needs retraining, and training-free caching methods from image generation say nothing about how their errors change the executed trajectory.

Proposed solution

Muninn runs only the cheap input stage of the denoiser at each step and scores how much that internal representation changed since the previous step. Split-conformal calibration on offline runs turns that score into a high-probability bound on the error from reusing a cached output, and analytic sampler coefficients convert it into a bound on final trajectory deviation. At run time the per-step costs are spent from a user-set deviation budget: the denoiser is reused while budget remains and recomputed otherwise, so the cached plan stays within a chosen distance of the full-compute plan with a chosen risk level.

Key insights

Diffusion samplers have a known analytic update, so the effect of a denoiser error at step t on the final trajectory can be bounded by sampler sensitivity coefficients; early steps can matter more than late ones
The change in a cheap intermediate representation between consecutive steps predicts how much a cached denoiser output would differ from a fresh one
Fixed step-skipping and fewer-step sampling spread error uniformly over diffusion time even though sampler sensitivity is highly non-uniform, which is why they lose task performance at the same latency
Budgeted reuse spends more denoiser calls in hard, cluttered situations and fewer in open space, without any extra training

Main results

D4RL AntMaze with Diffuser: latency 950 ms to 207 ms (4.6x) and denoiser calls per step 100 to 13.0, with normalized score 58.7 vs 58.3
At a 250 ms latency cap on AntMaze, Muninn scored 67.6 against 68.7 for the full model given 957 ms, vs 56.1 (FewSteps), 42.5 (FixedSkip) and 48.0 (ProbeThresh); violation probability 0.04
Hardware, 150 runs: ASV navigation with the BCOD teacher 60 ms to 24 ms latency (success 97.0% vs 96.8%); Crazyflie navigation with GC-Diffuser 75 ms to 30 ms (93.0% vs 92.8%); SO-ARM100 manipulation with Diffusion Policy 40 ms to 24 ms (82.0% vs 81.5%)
Cluttered 7-DoF arm planning (hard): MPD latency 230 ms to 121 ms with success 82.0% vs 81.2%; EDMP 160 ms to 107 ms with 88.0% vs 87.5%
Visuomotor policies: DP3 Pour 45 ms to 27 ms (98.0% vs 97.6% success); Meta-World pick-place 30 ms to 18 ms (92.0% vs 91.5%)
Measured probability that the cached plan exceeds the deviation tolerance stayed between 0.028 and 0.049 across the offline RL benchmarks in Table I

Limitations

The guarantee holds only for the chosen deviation metric and the calibration distribution; under distribution shift the exchangeability assumption can fail and risk control can degrade, so recalibration may be needed
The certificate bounds deviation from the full-compute planner, not task safety: if the base planner is unsafe, Muninn stays close to that unsafe behaviour
The union bound over timesteps can be conservative, reducing speed-up at very small risk levels or long diffusion horizons
It needs a cheap, informative probe of the model's internals, which may require engineering access for some architectures; changing the sampler, guidance or conditioning requires recalibration
How a deviation budget translates into closed-loop stability depends on replanning rate, tracker robustness and estimation noise, and is left to future work; the budget is fixed by the user rather than adapted to context

Field impact

Winner of the RSS 2026 Outstanding Student Paper Award and a finalist for the RSS 2026 Outstanding Paper Award. The authors release code, dataset and models.

Abstract

Diffusion-based trajectory planners can synthesize rich, multimodal robot motions, but their iterative denoising makes online planning and control prohibitively slow. Existing accelerations either modify the sampler or compress the network–sacrificing plan quality or requiring retraining without accounting for downstream control risk. We address the problem of making diffusion-based trajectory planners fast enough for real-time robot use without retraining the model or sacrificing trajectory quality, and in a way that works across diverse state-space diffusion architectures. Our key insight is that diffusion trajectory planners expose two signals we can exploit: a cheap probe of how their internal trajectory representation changes across steps, and analytic coefficients that describe how denoiser errors affect the sampler’s state update. By calibrating the first signal against the second on offline runs, we obtain a per-step score that upper-bounds how far the final trajectory can deviate when we reuse a cached denoiser output, and we treat this bound as an uncertainty budget that we can spend over the denoising process. Building on this insight, we present Muninn, a training-free caching wrapper that tracks this uncertainty budget during sampling and, at each diffusion step, chooses between reusing a cached denoiser output when the predicted deviation is small and recomputing the denoiser when it is not. Across standard benchmarks spanning offline RL planning (D4RL), configuration-space motion planning, and visuomotor diffusion policies, Muninn delivers up to 4.6× wall-clock speedups across several trajectory diffusion planners and diffusion policies by reducing denoiser evaluations, while preserving task performance and safety metrics. Muninn further certifies—at a user-chosen deviation tolerance and risk level—that cached rollouts remain within a specified distance of their full-compute counterparts, and we validate these gains in real-time closed-loop navigation and manipulation hardware deployments. Code, dataset and supplementary videos available at: https://github.com/gokulp01/Muninn.

Open Source Robotics