Every single work released this week is solving a problem that people actually hit when trying to deploy these systems outside lab conditions. We are no longer just proving these models work. We are fixing the stupid things that break them. That is the pattern nobody is talking about.

The camera calibration dead end

Every robotics team has wasted at least one engineer month on this. You train a perfect VLA policy in the lab, mount the robot on a cart, bump the camera 2 degrees, and success rate drops from 92% to 17%.

All existing view robust VLAs required you to feed calibrated extrinsics at inference time. Which means you still need calibration runs every time the camera moves. Which means you cannot just grab a random web cam, point it at the robot, and give it instructions.

CamVLA fixes this. They did the obvious thing that nobody did for three years: stop predicting actions directly in the robot base frame.

CamVLA architecture breakdown

Instead of outputting one action vector, the model outputs two completely separate values. First: what relative motion the end effector should make, described entirely from the camera's point of view. Second: the 6DoF transform that maps the camera frame to the robot base frame. A standard matrix multiply composes the two into the final action the robot executes.

No depth input. No stereo. No calibration. One RGB image. One text prompt. That is the entire input.

This decomposition works because it separates two unrelated problems. The policy learns what motion to make once, independent of camera position. A separate small head learns to read the camera position from the same image.

CamVLA benchmark results

This is the table that should make every robotics lab throw out their current VLA code:

Test condition	Standard VLA success rate	CamVLA success rate
Exact training camera view	89%	91%
15 degree pan offset	41%	87%
30 degree tilt offset	22%	82%
Camera moved 40cm sideways	17%	79%
Completely unknown new camera	11%	76%

The baseline does not just get worse. It breaks completely. CamVLA performance degrades gracefully. At 30 degrees tilt the baseline is effectively useless. CamVLA is still operating at 90% of its nominal performance.

Multiplayer world models stop treating people as scenery

This is the quiet shift that nobody noticed. Until this week every world model treated every agent except the ego as part of the background environment. That works fine for single player. It falls apart completely when multiple agents are interacting.

MIRA is the first world model that explicitly conditions on every agent's action stream separately. It does not try to predict what other players will do. It predicts what will happen given any combination of actions from any player. That is the difference between a video generator and a usable physics simulator.

Trained on 10,000 hours of Rocket League gameplay. 5B parameters. Runs 4 player real time at 20fps on one B200. Rollouts stay physically consistent for hours. Not seconds. Not minutes. Hours.

MIRA scaling properties

Prior world models always suffered exponential drift. The further you rolled out from the initial state, the faster they fell apart. MIRA does not do this.

The 5B model flatlines. FID does not drift. There is no collapse. This has never happened before for any generative world model.

The authors ran unguided rollouts overnight. They came back 12 hours later and the match was still running, following consistent physics, with no visible degradation.

UNIVERSE: stop separating imagination from action

UNIVERSE solves the other big failure mode of world models for driving. All prior architectures trained a video world model, then trained a separate action head on top. The action head never actually used any of the physics the video model learned. It just memorized training trajectories.

UNIVERSE runs both video prediction and trajectory planning inside the same DiT. They are not separate branches. They are the same tokens being denoised at the same time.

The clever trick is the visibility mask. During training the model sees both future video and trajectory. At test time you just mask out all the video tokens. You get all the benefit of the video supervision, with none of the runtime cost. 4.3x speedup. No accuracy loss.

Results on NAVSIM: 91.0 PDMS vs 89.6 for the best prior dual DiT design. Zero shot transfer to nuScenes and Bench2Drive works. No fine tuning required.

Small object detection finally gets frequency domain right

FSDC-DETR is the first detector that does not just throw away high frequency information during downsampling. Everyone knew this was a problem. Everyone accepted it as an unavoidable cost of multi scale processing.

They run two parallel branches the entire way through the encoder. One spatial. One frequency. They cross propagate information at every layer. They do downsampling separately in both domains before merging.

Results on VisDrone: +6.4 overall AP. +6.8 AP on small objects specifically. On AITODv2 it is +6.6 overall, +6.9 small. Those are not incremental gains. That is a generational jump for small object detection.

TimeThink: Video LLMs stop guessing

All existing video LLMs get the right answer for the wrong reason. They look at the whole video, guess the answer, then retroactively pick a timestamp that looks plausible.

TimeThink does the opposite. It rewards the model for first finding the correct time interval, then reasoning from that interval. Step by step reward for every part of the reasoning chain, not just the final answer.

Gains are consistent across every video reasoning benchmark. 7-12% improvement on temporal localization accuracy. Most importantly, when the model is wrong now it tells you which part of the video it was looking at. You can debug it.

What this all means

We just crossed an invisible line. None of these papers introduce a new capability that goes viral on twitter. All of them remove deployment blockers.

You can now mount a camera anywhere on a robot and it will work. You can run a physically consistent multiplayer world model on a single GPU. You can train a driving agent once and it will transfer to new cities. You can reliably detect 5 pixel objects in drone footage. You can ask a video model why it arrived at an answer.

None of this was true one month ago.

This is not the part of the cycle where we make impressive demos. This is the part where we build things that actually work.

Code and artifacts

All papers have accepted or pending ICML 2026 publication.

• CamVLA code and evaluation data: https://alibaba-damo-academy.github.io/CamVLA/
• MIRA playable demo, code and dataset: https://mira-wm.com
• FSDC-DETR: https://github.com/nevereverinsomnia/FSDC-DETR
• UNIVERSE and TimeThink code will be published before the conference starts.