The action generation layer of robot learning has quietly undergone a revolution in the past eighteen months. Diffusion Policy — Chi et al.’s 2023 paper that demonstrated score-based generative models could handle the multimodal, high-dimensional distributions that plague imitation learning — was a genuine breakthrough. But the field is now migrating to something faster, simpler to train, and better suited to the latency demands of real hardware: flow matching. Understanding why this transition is happening, and what it unlocks, matters whether you’re building manipulation systems or trying to read the next wave of robotics papers.

1️⃣ What Diffusion Policy actually solved — and where it strains

Before diffusion, robot policies trained with behavioral cloning collapsed in the face of multimodal demonstrations. If a human teleoperator sometimes grasps left, sometimes right, the policy averaged the two and grasped nowhere. Diffusion Policy solved this elegantly: by modelling the action distribution as a reverse-denoising process, the policy could represent sharp, distinct modes. The UNet and Transformer variants both worked. The problem is inference. A standard DDPM sampler requires 100 denoising steps; DDIM pushed that down to 10–25. At a control frequency of 10–30 Hz, spending 30–80 ms per action generation is expensive and, on edge hardware, often infeasible. Researchers worked around this with chunked action execution — predict a sequence of future actions, execute them open-loop, re-plan — but this trades responsiveness for throughput.

2️⃣ Flow matching: straight paths, fewer steps

Conditional Flow Matching (CFM), introduced by Lipman et al. in 2022 and operationalized in robotics primarily through rectified flow variants, learns a velocity field that transports samples from a simple prior (Gaussian noise) to the data distribution along approximately straight trajectories. In contrast to the curved, Langevin-diffusion paths that DDPM traces, these near-linear paths can be integrated accurately in as few as 4–8 function evaluations without quality collapse. The training objective is also cleaner: instead of matching a score function via denoising, you regress directly onto the velocity field connecting noise to data, which turns out to be lower-variance and faster to converge. On policy benchmarks, CFM-based policies match or exceed diffusion policy on task success while cutting inference cost by 3–5×.

3️⃣ π₀ as the proof of concept at scale

Physical Intelligence’s π₀ (Black et al., 2024) is the highest-profile demonstration of this shift. Built on a PaliGemma vision-language backbone and fine-tuned with a flow-matching action head, π₀ handles dexterous, contact-rich tasks — folding laundry, assembling boxes, bussing tables — that previous VLA architectures failed on. The flow matching head is central to why it works at real hardware speeds: generating a 50-step action chunk takes under 10 ms on the onboard compute they target. OpenVLA-OFT (open-source, University of Michigan, late 2024) followed a similar path, adding a flow-matching fine-tuning stage that meaningfully improved dexterous manipulation over the base VLA. The pattern is becoming a template.

4️⃣ What the next iteration looks like

The frontier now is consistency models applied to robot actions — single-step generation with quality competitive to multi-step flow matching, by distilling a trained flow model. Early results from labs including CMU and Stanford suggest this is viable and would push inference to sub-millisecond territory, finally decoupling action generation from the control loop entirely. Paired with hierarchical architectures that run slow language-level planning and fast flow-based reactive control at different timescales, you start to see the shape of a genuinely capable manipulation stack. The move from diffusion to flow matching isn’t hype-driven churn — it is the field resolving a real engineering tension between expressiveness and speed, and the papers landing in 2026 will assume this foundation.