ATSplat reports reconstructing a 3D scene from 12 images in under a second, then rendering new 512×960 views at 1,136 frames per second with 311,000 Gaussians. This morning also brings faster diffusion robot policies, VR control for miniature humanoids, language-guided grasping, long-video memory, identity-preserving video, full-song generation, multimodal compute controls, AI-license tracking, persistent multi-robot planning, and an image-editing benchmark.

What Changed Overnight

  • ArXiv’s July 23 announcement batch added new work on compact 3D reconstruction, robot control, video generation and captioning, multimodal inference, AI supply-chain licenses, persistent planning, music generation, and image-editing evaluation.
  • The New York State Technology and Engineering Educators’ Association has opened individual registration for its March 18–19, 2027 conference at SUNY New Paltz. Registration and payment close March 13.

Reconstructing And Controlling Spaces

ATSplat allocates more Gaussians where reconstruction is uncertain

Cho In, Jeonghwan Cho, Mijin Yoo, Gim Hee Lee, and Seon Joo Kim built a feed-forward 3D Gaussian Splatting system that starts with sparse 3D anchor tokens. An uncertainty score decides which tokens should expand, so the model can place more Gaussians in difficult regions instead of assigning a dense set to every input pixel.

On RealEstate10K and DL3DV, the authors report rendering quality at or above the tested feed-forward methods while reducing the Gaussian count by a factor greater than 5.7 compared with dense alternatives. With 12 input images at 512×960 resolution, ATSplat reconstructs a scene in under one second on one commercial GPU and renders new views at 1,136 FPS using 311,000 Gaussians. The abstract does not report performance on interactive or changing game scenes.

EVO searches for a faster diffusion-policy cache schedule

Diffusion robot policies repeatedly denoise an action sequence, which can make closed-loop control expensive. Siying Wang, Kangye Ji, Di Wang, and Fei Cheng treat cache refreshes across model blocks and denoising steps as one schedule, then use evolutionary search to find computations that can be reused without retraining the policy.

Across manipulation benchmarks, EVO reportedly preserved near-full task performance while accelerating action generation by up to 8.05 times. The lowest-compute schedule reduced the tested workload from 15.77G to 1.96G FLOPs. The search happens offline; the resulting schedule can be attached to a pretrained policy.

A miniature humanoid combines VR teleoperation with learned walking

Nicolas Kosanovic, Jordan Dowdy, and Jean Chagas Vaz developed a full-body telepresence stack for the ROBOTIS OP3. A human wearing a VR headset directs the upper body while a reinforcement-learning controller maintains balance and moves the legs.

The robot walked at up to 0.45 meters per second while moving its arms. In a cube-relocation demonstration, an expert operator moved an average of two 40-gram cubes within 10 minutes while the robot covered five meters. The experiment demonstrates remote manipulation, not autonomous task planning.

SeededGrasp separates language intent from grasp geometry

Yang Xu and collaborators ask a vision-language model to identify a seed point that matches the instruction, then pass that point to a smaller geometric grasp generator. This separates semantic choice from the final 3D hand or gripper pose and lets one system work across multiple robot forms.

The team is releasing a dataset with more than 2.5 million grasps in cluttered tabletop scenes. It reports 72% success in simulation and 78% in real-world experiments, above the tested baselines. The project page provides the code and data.

Video, Identity, And Music Generation

Self Gradient Forcing sends future-video errors back into context memory

Autoregressive video models often train on their own generated history, but later-frame losses do not necessarily teach earlier frames how to store useful context. Junhao Zhuang and collaborators add a second parallel training pass that recomputes the earlier context representation and lets future errors supervise that memory.

The authors report better subject identity, layout consistency, and temporal stability than Self Forcing across frame-wise and chunk-wise tests. A model trained on five-second windows generated rollouts lasting several minutes. The team says it will release the code and models.

PercepCap exposes object tracks and event times before writing a caption

Yifan Xu and collaborators make a video-captioning model produce a perception trace before its final description. The trace contains object trajectories and timed events, letting evaluators inspect whether an incorrect caption began with a spatial or temporal error.

Training combines supervised perceive-then-describe examples with reinforcement learning over both the trace and the caption. The authors report consistent gains over Qwen3-VL on direct captioning and caption-to-question-answer evaluation. Per-benchmark scores are absent from the published summary.

Vera targets identity drift in single- and multi-person video

Yulong Xu and collaborators assembled one million identity-aligned image-video pairs by matching the same person across clips. Vera emphasizes identity-relevant regions during training and controls how reference cues enter different layers of the diffusion transformer.

The paper reports stronger identity consistency, multi-person role binding, and motion naturalness while reducing subject swaps and over-copying from the reference image. The published summary gives no aggregate improvement number or public-release date.

A full-song system combines token planning with flow-based rendering

Junyu Dai and collaborators describe one framework for lyrics-to-song, instrumental, and cover-song generation. A hierarchical language model plans compressed audio tokens; a diffusion transformer renders a continuous audio representation; and a two-level melody module preserves melodic cues for covers.

The report compares several reward-based post-training methods and says the system is competitive on a multilingual automatic benchmark and the Artificial Analysis Music with Vocals leaderboard. It does not give an overall score in the abstract or describe a public listening test.

Builder Tools, Planning, And Evaluation

License labels often break between datasets, models, and applications

James Jewitt and collaborators traced 232,270 chains from Hugging Face datasets to models and then to GitHub applications. They define license laundering as either assigning a definitive downstream license after an undeclared upstream artifact or replacing one declared license category with another.

The study found that 62.3% of chains included at least one artifact with no declared license. Every license category carrying obligations survived end to end in fewer than 7% of chains, while the permissive category reached 95.1%. A license shown on a model or tool page therefore may not describe the obligations attached to its upstream training material.

SmartVL shares one compute budget across visual tokens and language layers

Pengcheng Wang and collaborators jointly control how many image tokens a multimodal model keeps and how much language-model computation it performs. A shared budget representation and differentiable latency estimate train the two controllers to divide resources according to input complexity.

The authors report better accuracy-efficiency tradeoffs than methods that prune visual tokens or language-model computation separately. The project page is public; the summary provides no single latency or accuracy figure for the gain.

Courteous planning leaves a shared environment easier for the next robot

Md Ridwan Hossain Talukder and collaborators study robots that receive tasks one at a time in a persistent home or restaurant. Their planner weighs the current route against predicted future costs for every robot, including states another robot may be unable to undo.

Across long task sequences, total cost fell 10.43% against a myopic planner and 4.03% against a selfish anticipatory planner in the two-robot home. In the three-robot restaurant, the respective reductions were 17.41% and 13.24%. The tests use persistent PDDL simulations rather than physical robots.

RS-RIE-Bench puts strict image-editing success at 24.28%

RS-RIE-Bench asks image editors to follow temporal, causal, and spatial rules in remote-sensing scenes. It separately evaluates whether the target region is plausible, untouched regions are preserved, and the result remains consistent with the sensor image.

Across eight open and closed models, the strongest achieved 24.28% overall accuracy when all requirements had to pass together. The mean relaxed joint success rate was 32.23%, and several open models approached zero in some causal or spatial categories. The domain is satellite imagery, but the preservation test resembles the demands placed on scene and asset editors.

New On The Events Calendar

Wonder News added the NYSTEEA Conference 2027 to the calendar. The technology and engineering education conference runs March 18–19 at SUNY New Paltz in New York. Educator and professional registration is open, with a January 15 early-bird cutoff and a March 13 final registration and payment deadline. The organizer says commercial exhibits will be present but has not published a separate exhibitor application window.

Watch Next

  • ATSplat code, memory use, and tests on dynamic scenes or creator-captured game environments.
  • Independent EVO measurements on control latency, robot hardware, and tasks outside the search benchmarks.
  • Public Self Gradient Forcing models and long-rollout comparisons using the same prompts, seeds, and duration.
  • Per-benchmark results and creator tests for PercepCap, Vera, and the full-song framework.
  • Repository tools that preserve upstream license evidence across datasets, models, generated assets, and shipped applications.
  • Physical persistent-environment tests for courteous planning and editor tests that preserve untargeted scene content.
  • NYSTEEA’s exhibitor application details and final 2027 program.

This article was written with assistance from Wonder Bricks AI Agent and edited by SunnyLabs.