Data Efficiency: All of that from very little data. Fine-tuned on half the real-world demos, Flex-π still beats every baseline trained on all of them. RoboTwin at 50 demos/task: 78.8%, against 41.9% for Fast-WAM, 31.4% for π₀.₅ and 17.2% for LingBot-VA. That's 1.9× the strongest baseline and 4.6× the weakest, at the demo budget where it matters most.

Generalization: we test all models under significant distribution shift, including novel distractors filling the workspace, and object types the policy has never handled. Flex-π holds 95.0% task completion on Put Plate and 70.0% on Sort Utensils — 2.5 and 5.0 points below its in-distribution scores. ManiFlow has 3D input and drops 32.5 and 22.5. On the unseen bag, π₀.₅ falls from 42.8% to 17.2%.

Dexterity: a deformable pouch. Unzip it, hold the mouth open, drop a pen in, re-grasp the pull and close it — six steps, each reachable only if the last worked, on an object with no stable shape. Task completion: Flex-π 70.0%, π₀.₅ 42.8%, ManiFlow 31.9%.

On 5 dexterous bimanual YAM tasks, Flex-π beats π₀.₅, Fast-WAM and ManiFlow on every single one — 2.3× the success rate of the strongest baseline. Even action-only, the cheapest policy here to run, is ahead of all of them. Precision: the robot repairs its own gripper. Eight stages in strict order, the tightest a 4mm bit into a 4.5mm socket — ±0.25mm. Flex-π finishes all eight in 11 of 20 rollouts. The best baseline manages it once. Imagine your robot can repair itself :)

Then we randomly drop visual streams during training, and force the model to generate the ones it never saw as input. We call this cross-modality forcing: imagine future 3D geometry with no pointmap given, future semantics from RGB and geometry alone, and so on. To do that, the backbone has to internalize each modality from the others, unifying appearance, geometry and semantics into one mutually predictive representation. One checkpoint comes out of that supporting any subset of streams, in and out. Action-only at 60ms, full joint generation at 193ms, anywhere in between. It's a runtime flag, not a retrain, so you choose your point on the speed–accuracy frontier after the model is trained.

Everyone's training WAMs now, and most of them predict one thing: future RGB latents. Video is a strong prior, but it's trained to reconstruct pixels, and carries no explicit signal for the accurate 3D geometry or object semantics that manipulation actually needs. Adding those usually costs something: an extra sensor, a new visual prior to train, or more latency at deployment. Here it doesn't have to. Geometry and semantics both come from the same RGB image via off-the-shelf DINO encoders, and the frozen Wan video VAE, trained only on RGB pixels, turns out to encode 3D pointmaps almost losslessly. So RGB, pointmaps and DINOv3 features all become token streams in one shared latent space, co-denoised with actions inside a Mixture-of-Transformers.

Are VLAs dead? No, and you don't have to choose between VLAs and WAMs. Introducing Flex-π: a multi-stream world-action model (WAM) that jointly predicts future RGB, 3D pointmaps and DINO semantics with actions in training, then deploys as a VLA, a full WAM, or anything in between — all from a single checkpoint. Website: flex-pi.github.io/

Hi, great question! Both Pi0 and ManiFlow are trained on the same data and evaluated in the same environment. The first result we show is comparing Pi0 with ManiFlow using only 50 demos, and ManiFlow shows better efficiency and generalization with 64.7% vs. 24.3%. Note the eval env is very challenging with randomized backgrounds and distractors, so it's understandable that Pi0 achieves a low success rate, which also highlights the efficiency of ManiFlow with 3D representation and better architecture/training objective. In the second experiment, we further scale the data from 50 demos to 500 demos. And you can see with more data, Pi0 increases from 24.3% to 94.0% with 500 demos, while our ManiFlow achieves 99.7% success eventually.

The real-robot experiment of ManiFlow (maniflow-policy.github.io/) is built on top of two great open-sourced teleoperation systems: - Bunny-VisionPro for bimanual dexterous hand with real-time retargeting using VisionPro (dingry.github.io/projects/bu…) - Open-TeleVision for humanoid teleoperation with immersive active visual feedback (robot-tv.github.io/) These two real-time teleop systems significantly accelerate the real-world deployment of ManiFlow across multiple robot embodiments! Please give them a try!