This is not another batch of incremental robotics papers.

Over 48 hours this week three independent groups dropped preprints that together resolve every major blocking limitation that has kept world action models stuck in demonstration benchmarks for the last two years. For anyone building physical robot systems, this is the point where the research stops being interesting and starts being usable.

The three broken things that just got fixed

For 18 months every WAM demo had exactly the same three failure modes. You have seen all of them.

1. Ask the robot to pick the screwdriver not the hammer. It will pick whichever object has higher average embedding similarity to the word screwdriver, which is random half the time.
2. Ask it to walk to the chair. It will generate perfect predicted video frames of arriving at the chair, then output motor commands that drive it directly into the wall.
3. Ask it to use pliers. It will grip the pliers by the jaw and hold them there forever.

Every single public demo you have watched was carefully staged to avoid these failures. This week all three are fixed.

MaskWAM: Language was always the wrong interface

Nobody building robots actually wanted language interfaces. We just did not have anything better.

MaskWAM does not abolish text prompts. It stops treating text as the primary source of spatial truth.

Prior WAMs ran text conditioning once at the start of the sequence, then forgot what you asked for. All spatial information leaked out over 12 transformer layers. The model would correctly predict that something moves, but had no idea which thing.

This paper does one very obvious thing that nobody tried for three years: it passes masks through every single stage of the model, as both input and output. You do not just prompt with a mask. The model is required to predict the mask of the target object for every future timestep alongside RGB and actions. This is implemented as a unified Mixture of Transformers where mask tokens share equal weight with visual and action tokens at every attention block.

On ambiguous LIBERO tasks where baseline WAMs scored 41% success, MaskWAM hit 78%. On identical tasks where the prompt was just a 1 bit mask instead of text, it hit 92%.

That is not a small improvement. That is crossing the threshold from works sometimes to works reliably.

World action models see pixels. They do not read words. Stop trying to translate human language into robot coordinates. Just draw a box around the thing you want moved.

The mask supervision trick that everyone will copy

The real insight in MaskWAM is not using masks as input. It is using mask prediction as an auxiliary loss during training.

Prior WAMs only trained on RGB reconstruction loss. This meant the model got exactly the same reward for moving the correct object, moving the wrong object, or just shifting the lighting in the scene. All three produced equally low pixel error. No penalty was applied for manipulating the wrong thing.

MaskWAM adds an extra head that predicts segmentation masks for every future frame. This loss is weighted equally with RGB and action loss.

This changes everything. Now the model gets penalized explicitly if the wrong object moves. It does not get to cheat by changing background pixels. It has to track exactly the object you specified, for every step of the trajectory.

The authors tested this by adding only the mask prediction loss to existing open source WAM implementations, with no other architecture changes. It improved baseline performance by 27% across every benchmark. You can add this to your existing model tomorrow.

NavWAM: Kill the planner

Until this paper every world action model for navigation had a dirty secret. None of them actually output actions.

They would predict 32 possible future trajectories. Then they would run an external Cross Entropy Method planner over those predictions to pick the best one. Then they would execute the first step. Then repeat.

This was always a scam. The world model was not controlling the robot. The planner was. The world model was just a very expensive lookup table.

NavWAM removes this entirely. It trains the diffusion transformer to output action chunks directly alongside predicted observations and goal progress value. There is no search step. There is no external planner. Forward pass the model once, get the next 8 actions, execute them.

On real robot navigation tests, planning based WAMs achieved 62% success rate. NavWAM achieved 79%. It also ran 11x faster.

This is the part that matters for deployment. You do not need an A100 on the robot. This runs at 12hz on a Jetson Orin.

Why NavWAM works when everyone else failed

The critical change is that NavWAM does not train prediction first then attach a policy head later.

All prior work followed the obvious pipeline: pretrain the world model on observation prediction, freeze it, then train a policy on top. This never worked. The world model learned features that were very good for generating pretty video, and completely useless for choosing actions.

NavWAM trains all three heads, observation, value, action, from the very first training step. Gradients flow backwards through all three at every update. The model never gets to learn features that do not help choose actions.

This is a general principle, not just for navigation. Every world action model built from now on will be trained this way.

Mana: We stopped treating hands like robots

While everyone was arguing about WAM architectures, the Mana team went and solved articulated tool manipulation.

This is the problem everyone said was 5 years away. Opening scissors. Turning a wrench. Operating a lighter. Everything that makes a hand useful instead of just a clamp.

They did not do it with a bigger transformer. They did not do it with more real world data.

They copied character animators.

Mana works like this:

1. For a given tool, you click once on the part that moves, once on the part you hold. That is all the human input required. Total annotation time is less than one minute per tool.
2. The system procedurally generates 100,000 valid grasp keyframes. Not trajectories. Just start and end poses.
3. A sampling based motion planner fills in the space between keyframes.
4. A small 12M parameter RL policy fine tunes the trajectory to tolerate contact noise and simulation error.

That is it. No foundation model. No billion parameters. Zero shot sim to real transfer worked on every tool they tested.

Success rate for opening a pair of pliers one handed was 87% on a real Shadow Hand. Prior published work on this exact task topped out at 19%.

The unstated lesson from Mana

This paper is an open insult to every robotics team that spent the last three years scaling demonstration datasets.

You do not need 1000 human hours of teleoperating pliers. You need to know that you hold the handles and squeeze. That is one bit of information. Everything else can be generated.

The entire field has been wasting time trying to imitate human motion when we should have been defining the constraints and letting the computer fill in the motion. Animators figured this out in 1995. We just got here.

What this means for production robots

None of these papers are perfect. All have limitations. None of them will work unmodified on your hardware.

That does not matter.

Until last week, every production robot deployment was either hard coded, or trained on exactly one task for exactly one environment. General purpose policies were demo bait.

That is over. Right now you can take these three architectures, combine them, and build a robot arm that you can point at an arbitrary object, give it a simple goal, and it will succeed most of the time. No fine tuning. No task specific training.

We are no longer waiting for a breakthrough. We are waiting for people to implement what has already been published.

The quiet end of the foundation model robot era

For three years the dominant approach was: take a general purpose vision language model, attach a robot arm, and hope.

All three papers this week reject that approach completely. None use LLMs. None use frozen foundation model backbones. All build small, specialized models trained exclusively for physical interaction.

They win by a factor of two.

This is the turning point. The era of trying to repurpose internet models for the physical world is ending. The next generation of robot models will be built for the physical world from first principles.

These three papers are the start of that generation.

---

References

1. Mana: Dexterous Manipulation of Articulated Tools. http://arxiv.org/abs/2606.13677v1
2. MaskWAM: Unifying Mask Prompting and Prediction for World-Action Models. http://arxiv.org/abs/2606.13515v1
3. NavWAM: A Navigation World Action Model for Goal-Conditioned Visual Navigation. http://arxiv.org/abs/2606.13494v1