What WAM Actually Optimizes
The hottest term in the 2026 world-model race is WAM — World Action Model. GigaWorld-Policy pushes inference down to 85 milliseconds on an RTX 4090. Coowheel's DAWN architecture lets world prediction and action generation "negotiate" in latent space. Xingyuanzhi's ω-EVA makes robots rehearse consequences before they act. Everyone is trying to prove the same thing: robots should "think before they move."
But action was never the point of work. Ask a robot to fit part A into slot B, and the task is not "rotate the wrist, close the gripper, apply 5 newtons" — those are three actions. The task is "it fits, it's tightened, and nothing is damaged." WAM stares at the question "is the next move correct?" and ignores "will the whole job get done?"
That blind spot is structural. A WAM is an action optimizer, not a task finisher. Picture an obsessively disciplined marathon runner — perfect pacing, breathing, arm swing — who has never looked at the course map. Flawless single steps, poor chance of finishing.
The multiplication effect does the rest of the damage. In a 10-step task, a 95% success rate per step compounds down to roughly 60% overall. Even 99% per step yields only about 90% end to end. That is why VLA and WAM demos keep collapsing on show floors: every step looks stable in isolation; chained together, they fall apart.
Why "Task Completion" Becomes the First-Principles Metric
Enter WTM — World Task Model. Instead of scoring each action in isolation, a WTM explicitly tracks task progress, sub-goal completion, and final success probability as the core supervision signal. Every decision is weighed against one question: does it move the job closer to done?
The difference from WAM shows up in three places:
- Optimization target. WAM maximizes the probability of the next correct action; WTM maximizes the probability of the whole trajectory succeeding.
- Error handling. In WAM, per-step errors accumulate independently and compound by multiplication. In WTM, errors get absorbed in task-progress space — the model notices progress stalled and corrects, instead of letting drift ride to the end of the task.
- Data honesty. "95% single-step accuracy" sounds impressive on a slide and gets embarrassed on a factory floor. WTM's metric is task success rate — a number a buyer can read and trust.
The clearest analogy: WAM is a craftsman who obsesses over each carve; WTM is the project manager who signs off on whether the house is habitable. Customers pay for a habitable house, not for evenly carved wood.
Who Is Closest to WTM
Ranked by how close each player is to a true task-level model:
- Physical Intelligence — closest VLA player. π0.5 introduced high-level subtask prediction; π0.7 added visual subgoal images and episode metadata. That is a task-progress and sub-goal representation squeezed inside a VLA shell. In an untrained two-arm shirt-folding test, π0.7 hit 80.0% zero-shot — nearly matching a human teleoperator's 80.6%. PI is one thin window away from WTM: it still treats subtasks as conditioning inputs rather than joint training signals.
- Xingyuanzhi — WTM's skeleton, still WAM. ω-EVA's Envision-Verify-Act loop brings world prediction into the action-decision loop, and learning from failure data is close to task-level feedback. But its correction loop stays inside a local "state + candidate action + imagined outcome" triple, with no explicit task graph or progress head. It is 1.2B params, LIBERO 98.6% — single-task territory.
- Coowheel — strongest WAM in continuous control. DAWN's recursive world-predictor/action-denoiser correction is impressive (89.1 PDMS on NAVSIM), but driving is a continuous flow with no clean sub-task boundaries. It is a parallel universe to manipulation tasks.
- GigaWorld — pure WAM engineering. It optimizes deployment over reasoning, even disabling future-video generation at inference to hit 85 ms. That is a deliberate compromise away from "world" and from "task" — the farthest from WTM by architecture.
- OAKMU — the wildcard. It openly skips VLA and world models, betting on instinct-driven, touch-closed-loop execution with near-zero error per step — pushing the multiplication product toward 1. Its changeover time of 6–40 minutes (vs. weeks for traditional lines) is already task-level evidence, even without a world model.
What Actually Forces the Shift: The Buyer's Ledger
The catalyst is not a paper — it is the buyer's ledger. KPMG's WAIC report says investment evaluation has left the "bigger parameters, prettier demos" era and now centers on task completion in real scenes, generalization to unfamiliar environments, and unit deployment cost. Diligence now looks at production lines and repeat orders, not slide decks.
Factory RFPs are blunter: "Can this line's yield go up 5 points? How long to change an SKU — a week, or minutes? What is the payback on this CAPEX?" Leading 3C makers have made task completion and changeover efficiency veto conditions in tenders. With 90 billion RMB poured into embodied AI in the first half of 2026 (5x year over year), the demo era is over — someone must start talking about payback. Our recent look at how AI startups survive a funding winter showed the same rule: capital eventually demands results, not runway.
This is why the deadline is roughly eight months. Regulators want humanoid robots producing normal, sustained output on real lines by year-end; 3C tenders are already voting on task completion; capital needs to see a return. Stack those clocks together and the window for a task-level answer lands in early 2027. The same principle that forces multi-agent systems to add generate-run-check-repair loops applies here: reliability only counts at the level of the finished result.
What Practitioners Should Do Now
- Track task success rate, not single-step accuracy. If a model's benchmark is "LIBERO 99%" but nobody can tell you the end-to-end pass rate on a 50-step batch, the metric is hiding the failure mode.
- Watch for the progress head. The telltale sign of WTM-ization is an explicit task-progress or sub-goal regression head entering the joint training objective.
- Price against changeover, not demo elegance. Minutes-scale SKU changeover is the metric buyers are actually writing into RFPs.
- Accept that WTM is an outcome, not a specific architecture. World-model, VLA-shell, or instinct-driven — if it reliably finishes the task, it is WTM in effect.
The toy era of embodied AI is over; the tool era has begun. The question is no longer how fast a model runs, but whether it can tick the "task completed" box on a real production line. WTM will not be born in a lab. It will be forced into existence by buyers who are done waiting.
