Predict where your hands will be two seconds from now and dexterous manipulation gets a lot easier. Four papers from this cycle make that bet from four different directions. P3-WAM folds perception, prediction, and planning into one set of weights (arXiv:2610.22363). Whole-Body Dexterity turns the torso and forearms into contact surfaces instead of trusting fingers alone (arXiv:2610.22184). Velox makes long video rollouts cheap enough to drive an interactive simulator (arXiv:2610.22311). DSP gets a robot to generate its own contact-rich training data by playing (arXiv:2610.22340).

Quick take: the policy stopped being the hard part. What's hard now is generating contact-rich experience and keeping the world model fast enough to matter inside a control loop.

The constraint moved from hands to hours

Five years ago, dexterous manipulation was a hardware problem. Multifingered hands cost more than a car and broke every few weeks of use. Today you can buy a Unitree G1 with a working hand and a usable SDK for less than a used sedan. The bottleneck moved. What's scarce is contact-rich experience: the hours of a real hand sliding, bracing, and dropping things that a policy needs before it generalizes past its training set.

DSP's headline number makes the shift concrete. It reaches reinforcement-learning-comparable success on in-hand reorientation using 490 times less interaction data. If your RL baseline needed two weeks of continuous robot time, you're now looking at roughly forty minutes of seed play plus autonomous rollout. The bottleneck is no longer wall-clock robot hours. It's the quality of the seed play and the coverage of the world model trained on top of it.

P3-WAM: three modes, one set of weights

The representation is the interesting part. P3-WAM doesn't feed raw joint angles to a video model. It builds a compact, structured 3D action representation (articulated hand motion plus camera trajectory) that lives as tokens next to video patches. The mixture-of-transformers sits on shared base weights with modality-specific adapters, so visual and action tokens attend to each other in both directions.

Randomized modality visibility is what lets one checkpoint serve three jobs. Sometimes video is input and action is the target. Sometimes the reverse. Sometimes both are conditioning context.

The practical payoff: one model covers hand reconstruction, action-conditioned video prediction, and goal-directed planning. You're not chaining a pose estimator, a dynamics model, and a planner. That's three fewer coordinate-frame conversion bugs sitting between you and a working demo. Unified training beats single-mode training on all three capabilities across four datasets, which usually means the shared trunk is doing real work rather than just regularizing the task heads.

Whole-body contact is a different manipulation problem

Whole-Body Dexterity asks a fair question. Why do robots only grip with their fingers? Humans brace boxes against the forearm, pin objects between upper arm and ribs, and slide heavy things along the torso. WBD builds a dataset of feasible whole-body grasps in sim, trains reinforcement-learning experts on an any-to-any repositioning objective, then distills them into a single policy driven by proprioception and object geometry.

The real-world numbers are the story.

A 90% completion rate on a real G1 against 18% for direct retargeting and 4% for a generalist whole-body controller. If you've shipped a sim-trained controller, you know how unusual a five-fold gap over retargeting is. The distilled policy repositions a basketball and a cardboard box around the torso, then runs a long-horizon task that stows a box under one arm, carries it from a shelf to a table, and places it. Coarse motion references come from a human teleoperator, so the controller is interactive rather than fully autonomous. That's a reasonable division of labor for now: the human picks the goal, the policy handles contact.

Four papers, side by side

Work	Core mechanism	Reported result	Real cost
P3-WAM	Mixture-of-transformers over video and 3D action tokens	Beats single-mode training on hand reconstruction and action-conditioned video prediction across four datasets	Paired video and articulated hand pose
Whole-Body Dexterity	Sim-generated grasp dataset plus RL experts distilled to one policy	90% of real G1 trials	Sim-to-real pipeline plus teleoperator for coarse motion refs
Velox	Training-free frustum-slab attention router	4.36x attention, 2.43x denoising at chunk 48	Streaming depth, H200-class GPU
DSP	Play policy feeds a world model, which drives MPC	490x less data than RL, 69.6% over hindsight behavior cloning	Minutes of human play, hours of autonomous rollout

Velox and the caching tax on long rollouts

Chunked video diffusion has a dirty secret. The KV cache grows with every chunk, so chunk 10 is fast and chunk 50 crawls. For an interactive world model, that's fatal. The whole point is responding to your action within a frame or two.

Velox attacks the cache directly. It represents historical blocks as camera frustum slabs with depth intervals, obtains current-block depth through recency-priority multiwarp, and uses slab overlap to softly modulate each layer's block importance. Uncovered regions keep the model's own routing without geometric bias. Token-block granularity matches FlashAttention tiling, so it slots into existing kernels without retraining.

At chunk 48, Velox speeds up attention by 4.36x and denoising by 2.43x, landing within 0.15 API-free average points of dense LingBot-Fast-Cam across the full 158-case WBench navigation split. Frustum slab construction runs asynchronously with denoising on the same H200.

That 0.15-point regression is the number that matters. A training-free router that costs you a fifth of a point to buy a four-times speedup at long horizons is close to free. What I'd want to see next is how the speedup curve behaves past chunk 100, because the appeal is specifically that the term growing with rollout length stops growing.

Play as a data engine

DSP's structure is worth restating because it's the cleanest version of an idea that has been circulating for a while. Clone a play policy from minutes of human play. Deploy it autonomously to collect hours of interaction data. Train a world model on the collected data. Then use the same play policy to propose candidate action sequences for model-predictive control under the world model.

The same policy that generated the data becomes the sampler at inference. Economical. It also means the world model's coverage is bounded by the play policy's exploration, which is a real constraint on any task the policy won't stumble into. They also distill a closed-loop, goal-conditioned policy through self-play inside the world model to cut inference cost.

On underactuated pushing and any-to-any in-hand pose reorientation with a multifingered hand, DSP beats hindsight behavior cloning by 69.6% in real-world success rate. No task-specific demonstrations. No sim-to-real transfer step. That last part is the one to pay attention to.

Key numbers, in practical terms. 490x fewer interaction data than RL for DSP's in-hand reorientation result, which turns a two-week data campaign into a forty-minute one. Also: 90% trial completion for the distilled whole-body policy on a real G1, against 18% for direct retargeting and 4% for a generalist controller. Velox's 4.36x attention speedup at chunk 48 arrives with 0.15-point regression on the 158-case WBench split. DSP's 69.6% edge over hindsight behavior cloning holds on real hardware.

What I've seen running rollouts

I spent a week trying to keep a chunked video world model coherent past thirty chunks, and the failure mode wasn't visual drift. It was contact. The hand looked perfect. The box looked perfect. Then the model quietly predicted the box passing through the fingers. Latency was fine, PSNR was fine, the demo looked fine. MPC built on top of that model goes nowhere.

That experience reshaped how I read these papers. Visual fidelity metrics reward smoothness, and smoothness is exactly what a contact-rich prediction should not have. A world model that renders a blurry hand making solid contact beats one that renders a crisp hand ghosting through an object every single time.

The other pattern I keep running into across teams using this class of method: the caching policy and the rollout policy have to be co-designed. If the simulator runs at 30 Hz and the world model updates at 4 Hz, the planner is always reasoning about a state that's already gone. Velox-style routing helps, but only when the control loop can tolerate the staleness. Nobody mentions this in the abstracts, and it's usually what breaks the demo.

What trips people up

Evaluating a world model on visual metrics. PSNR, FVD, and LPIPS all reward temporal smoothness, which is the opposite of what contact looks like. Pick a contact-specific metric (penetration depth, object displacement error after first contact) before you tune anything else.

Training the play policy and the world model on the same distribution. DSP works because the play policy explores past its own training data during autonomous rollout. Freeze that distribution too early and the world model just learns the mean, then MPC inherits every blind spot.

Skipping the feasible-grasp dataset for whole-body tasks. Direct retargeting dropped to 18% on the G1 while the distilled policy hit 90%. That gap is the dataset doing real work. You can't RL your way out of a contact configuration that isn't physically reachable.

Letting chunked KV caches grow unbounded in something you're calling interactive. Past roughly forty chunks, attention dominates denoising time. If your latency curve bends upward, you have a video player with extra steps, not a simulator.

Treating 3D hand reconstruction accuracy as a proxy for manipulation quality. P3-WAM improves both jointly, but they're separate capabilities. A millimeter-accurate hand pose is useless if the planner can't sequence it into a grasp.

One thing to remember

Across all four papers, the world model isn't a research artifact sitting next to the policy. It's inside the control loop: as the data generator (DSP), the planner (P3-WAM), the interaction surface (Velox), or the contact model (WBD). Once you accept that placement, the design constraints flip. Latency matters more than rendering fidelity. Contact accuracy matters more than pixel sharpness. Every millisecond of cache growth compounds through the planner instead of disappearing into a batch.

The bottom line

Have a dexterous hand and no task demos? Don't reach for RL first. DSP hits comparable success with 490 times less interaction data, so the sensible opening move is twenty minutes of human play plus an autonomous world-model training loop, then MPC on top.

Your world model's latency problem is probably the KV cache, not the denoiser. Velox buys 4.36x on attention at chunk 48 with zero retraining and a 0.15-point WBench regression. If rollouts past forty chunks aren't getting faster, look at the router before you look at quantization.

Skip direct retargeting for whole-body tasks. It finishes 18% of trials against 90% for a distilled policy on a real G1. Budget for sim-generated grasp datasets and RL expert distillation, and treat your teleoperator as a goal-setting interface rather than a temporary crutch.