We finally stopped chasing bigger models for robotics

For three years every embodied AI announcement followed the exact same script. Take a general purpose MLLM, fine tune it on 1000 robot demonstrations, make it stack three cups in a lab, write a press release. None of them ever left the lab.

This week that ended. All six papers dropped on arXiv June 9th solve problems that nobody was talking about, but every single robot deployment was dying on. None of them use models larger than 7B parameters. Most improve existing policies without retraining the base.

This is not incremental progress. This is the stack of fixes that will let embodied AI leave the demo stage.

The single dumbest limitation of every VLA is now fixed

Every Vision-Language-Action policy ever released ran at exactly one speed.

Nobody talked about this. If you watched any robot demo video you already noticed it. Every motion happened at the exact same glacial, uncanny pace. The robot moved exactly the same speed when reaching across an empty table as it did when threading a USB cable.

You could not make it go faster. You could not make it go slower. You could only swap out the entire policy for a different one trained at a different fixed speed. All prior work on VLA acceleration only shifted that fixed speed upwards. Nobody had even attempted to build deceleration.

TempoVLA fixes this completely.

The authors noticed something that every single person working on VLAs had walked right past: the magnitude of the predicted action delta is exactly what controls speed. There is no hidden speed parameter. The policy already outputs everything required. All you needed to do was condition it.

They did two things. First they built Variable-Speed Trajectory Augmentation, which takes every existing demonstration and retimes it to every possible speed between 0.2x and 3.5x while preserving motion shape. This is just data preprocessing. It does not require new demonstrations. It does not require RL. You can run this on every existing VLA training dataset tonight.

Second they added a single scalar speed condition token to the policy input. That is the entire model change. One token.

In testing TempoVLA hits requested speed within 2.1% error across the full range. It can accelerate mid task. It will slow down automatically when approaching contact. And as an accidental side effect, the augmentation alone improved baseline 1x task success rate by 12.7%.

This is not a clever new architecture. This is someone finally looking at the thing everyone was using and noticing the obvious thing nobody saw. Every VLA shipping 12 months from now will have this.

Robots can finally see parts, not just whole objects

PAR3D fixes the second silent failure mode of every embodied system.

All existing 3D MLLMs are object centric. They can tell you there is a door in the scene. They cannot tell you there is a handle on the left side of that door. They can see a chair. They cannot see the armrest you need to grab to move it.

This is why every robot demo only uses perfectly clean unobstructed objects. This is why every real world deployment fails the second something is broken, partially occluded, or just not a standard object.

PAR3D adds hierarchical part representation into the 3D tokenization pipeline. It does not require retraining the entire MLLM. It injects part level queries during attention, and grounds them against a new 1.2M annotation ScenePart dataset.

On part level referring segmentation it beats all prior 3D MLLMs by 41.2%. It still matches or outperforms baseline models on all standard object level tasks.

Most importantly: this works on unmodified real world point clouds. You do not need CAD models. You do not need pre-scanned objects. If there is a broken latch on a cabinet, PAR3D will see the latch. It will understand you are asking to open it by the broken part.

This was the single largest capability gap for general manipulation. It is now closed.

We will never run out of simulation scenes again

HomeWorld solves the simulation bottleneck.

Everyone working in embodied AI knows this dirty secret: all of our models are overfit to the same 1200 simulation scenes. Every benchmark runs on the same houses. Every policy is trained on variants of the same furniture. The moment you deploy them anywhere else they fall apart.

We could not generate good new scenes. Prior generation methods would produce rooms where chairs floated through tables, kitchens had no sinks, doors opened into walls. You could not run physics on them. You could not run a robot in them.

HomeWorld decomposes scene generation into three separate staged models. First an LLM trained on 300,000 real residential floorplans generates valid global layout. Then a diffusion model places furniture. Then a VLM refiner iteratively fixes placement errors.

The output scenes are physically valid, fully interactive, and indistinguishable from real houses in blind user testing. The authors are releasing 5000 prebuilt scenes, and the full pipeline that can generate an unlimited number of unique, consistent whole homes.

This is not just for interior design. This removes the single largest data bottleneck for indoor robot training. Within 6 months every major robotics lab will be training on 10 million generated scenes.

Multimodal attention was broken this whole time

GRAMformer fixes a fundamental flaw that every multimodal model has been carrying for 4 years.

All existing cross attention only models pairwise interactions. When you feed vision, audio, depth, and joint state into a multimodal model, the attention mechanism can only compare query to vision, query to audio, query to depth. It can never ask: what is the combined meaning of this depth reading, this camera pixel, and this torque reading together.

Nobody noticed this was missing. Everyone just kept stacking more attention heads.

Volumetric Multimodal cross Attention calculates attention score as the signed volume of the simplex spanned by the query and all key vectors across modalities. It natively captures 3-way, 4-way, any order interactions between any number of input modalities.

For 4 input modalities GRAMformer achieves 18% higher task accuracy while using 42% fewer parameters and running 27% faster.

This is not a robotics specific result. This will change every multimodal architecture built from this point forward.

End to end driving just crossed the deployment threshold

CLEAR is the first end to end driving model that is fast enough to actually use.

For two years everyone knew diffusion based driving plans produced better, more human like trajectories. Everyone also knew that 12 step denoising at 10hz was never going to ship in a car. Safety critical systems require sub 100ms end to end latency.

CLEAR throws out the iterative denoising loop entirely. It does a single conditional drift step in VAE latent space. It uses a tiny 0.8B Qwen fine tuned on driving QA to adjust the planning coefficient dynamically per scene.

On NAVSIM v1 it hits 93.7 PDMS, the highest score ever published. It runs end to end in 72ms on an Orin NX. That is production ready. That will run on hardware already installed in cars on the road today.

Nobody will be running iterative diffusion for driving 12 months from now.

Disaster response is the first real embodied AI use case

DisasterBench is the first benchmark that does not test for demo performance. It tests for performance when everything is broken.

Every existing multimodal benchmark asks: describe this image. DisasterBench asks: this roof is buckling, how long until it collapses, which way will it fall, and where should you land the UAV.

It tests causal reasoning, failure propagation, damage estimation, and action selection. All on noisy low altitude UAV footage, all under edge compute constraints.

The accompanying DisasterVL 2B model matches GPT-4o reasoning accuracy on this benchmark while running at 28fps on a Jetson Orin.

This will be deployed before the end of this year. There are fire departments already testing this. This will not be a demo. This will save lives.

What happens next

None of these papers got any press. None of them announced a new one trillion parameter model. None of them have flashy demo videos going viral on twitter.

That is exactly why they matter.

Embodied AI did not need a smarter model. It needed to stop being bad at all the boring obvious things. It needed to be able to go fast and slow. It needed to see door handles. It needed enough simulation data to stop overfitting. It needed to run fast enough.

All of that is now fixed.

We are not 10 years away from useful general robots. We are 18 months away.

The only thing that was stopping us was that nobody had bothered to fix these problems. Everyone was too busy building bigger demo models.

This week that changed.

References

1. TempoVLA: Learning Speed-Controllable Vision-Language-Action Policies http://arxiv.org/abs/2606.06491v1
2. PAR3D: A Unified 3D-MLLM with Part-Aware Representation for Scene Understanding http://arxiv.org/abs/2606.06485v1
3. HomeWorld: A Unified Floorplan-to-Furnished Framework for Generating Controllable, Densely Interactive Whole-Home Scenes http://arxiv.org/abs/2606.06390v1
4. GRAMformer: Any-Order Modality Interactions via Volumetric Multimodal Cross-Attention http://arxiv.org/abs/2606.06249v1
5. CLEAR: Cognition and Latent Evaluation for Adaptive Routing in End-to-End Autonomous Driving http://arxiv.org/abs/2606.06219v1
6. DisasterBench: A Multimodal Benchmark for UAV-Based Disaster Response in Complex Environments http://arxiv.org/abs/2606.06217v1