Research map / Image generation / Autoregressive and latent generation

Efficient Diffusion Transformer Policies with Mixture of Expert Denoisers for Multitask Learning

ICLR 2025 · Image generation › Autoregressive and latent generation · approach: Latent diffusion with learned tokenizers · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MoDE, a sparse mixture-of-experts diffusion transformer policy that routes tokens by noise level, with noise-conditioned self-attention and expert caching, achieving SOTA on CALVIN and LIBERO with lower inference FLOPs.

Problem

Diffusion transformer policies for goal-conditioned imitation learning are computationally expensive as they grow in size and require many denoising steps, limiting real-time robotics use.

Approach

Use a mixture-of-experts transformer whose router is conditioned only on the noise level, so the expert path at each denoising step is deterministic and can be precomputed and fused into a single MLP (expert caching), combined with noise-conditioned self-attention and a noise token.

Key results

• Pretrained MoDE achieves 4.01 average rollout length on CALVIN ABC→D, versus 3.06 for GR-1 and 1.35 for the CNN diffusion policy.
• MoDE reaches 0.92 average success on LIBERO-10 and claims the first over-90% success on that suite; it uses 1.53 GFLOPS per action vs 27.5 for GR-1 and 16x fewer FLOPs than dense transformer at batch size 512 (361 vs 5,772 GFLOPS).
• On SIMPLER, MoDE averages 26.30% success vs 23.70% for OpenVLA and 17.75% for Octo.

Measured effects reported in the paper

Surprising findings

• Using only the top-1 expert lowers LIBERO-10 average success only marginally (0.92 to 0.91), and a shared single expert still achieves 0.90.
• Increasing experts beyond 4 degrades CALVIN ABCD performance.

Limitations

• Slightly higher standard deviation than baselines, attributed to router initialization sensitivity.
• Expert collapse observed in some experiments where only a subset of experts is used.
• Pretrained weights and preprocessing code are promised but the data release is not stated.

Experimental setup

Datasets: CALVIN LIBERO-10 LIBERO-90 Open-X-Embodiment subset (BC-Z, BRIDGE, CMU Play-Fusion, Google Fractal, DOBB-E) Relay Kitchen Block Push

Benchmarks: CALVIN ABCD→D CALVIN ABC→D LIBERO-10 LIBERO-90 SIMPLER Relay Kitchen Block Push

Metrics: average sequence length success rate FLOPs inference time (ms) rank

Baselines: Diffusion Policy-CNN Diffusion Policy-Transformer QueST RoboFlamingo GR-1 SuSIE OpenVLA Octo

Builds on: Diffusion Policy EDM Switch Transformer DDIM

Closest papers by meaning

• Two-Steps Diffusion Policy for Robotic Manipulation via Genetic Denoising (NeurIPS 2025 · similarity 0.90)
• Remix-DiT: Mixing Diffusion Transformers for Multi-Expert Denoising (NeurIPS 2024 · similarity 0.89)
• Prediction with Action: Visual Policy Learning via Joint Denoising Process (NeurIPS 2024 · similarity 0.89)
• Denoising Task Routing for Diffusion Models (ICLR 2024 · similarity 0.89)
• Latent Diffusion Planning for Imitation Learning (ICML 2025 · similarity 0.88)
• HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.88)