Real-world scenes are highly repetitive 🧱🌿: bricks, blades of grass, leaves, and many other structures appear again and again. At #NeurIPS2026, we present SCION, which learns a compact vocabulary of reusable 3D primitives and instantiates them throughout a scene, capturing its underlying compositional structure. This brings learned photorealistic 3D Scenes down to the ~1 MB range 🤏 while introducing natural structures for editing and animation 🎬. Great work by @williamkoch314, @amogh7joshi, @cvachha, and Cheng Zheng. You'll see the authors in Sydney this December! Project Page: light.princeton.edu/SCION Paper: arxiv.org/abs/2610.02322

Super excited to share this unedited fully autonomous truck ride! We're seeing strong zero-shot planning capabilities emerge from self- and mixed-play RL training at scale in complex situations. Our stack has not been trained on scenarios from this unseen route, and the generalization capabilities are tremendous! Proud of the @torc_robotics team for their work on this frontier architecture for autonomous driving.

ScenarioControl (ECCV 2026) enables end-to-end controllable generative simulation 🚗✨for autonomous driving, allowing users to generate diverse traffic scenarios with fine-grained actor control while preserving realistic scene dynamics. We have released the code, and we’re looking forward to seeing what the community builds with it! 🔗 Check out our project page for the paper & code: princeton-computational-imag… 👏 Thanks to the first authors @Lili_Gao_ , @Yanbo_Xu_ , @WilliamKoc61157

End-to-end Self-play from Pixels! Excited to share Gigapixel: a high-throughput driving simulator that enables scalable self-play training of end-to-end driving models! Super fun work from the amazing @Luke22R ! We extended PufferDrive with a GPU-accelerated perspective renderer, enabling self-play training from pixel observations at 50k SPS on 1 GPU, with throughput scaling linearly with the number of GPUs. We also introduce self-play DAgger training, a sample-efficient alternative to self-play RL that produces robust driving policies in ~3000x fewer steps. Planners trained in Gigapixel transfer to photorealistic observations via lightweight perception adaptation, achieving SOTA on closed-loop driving benchmarks, all without human trajectory supervision. Moreover, scaling self-play yields proportional gains in policy performance, establishing it as a scalable strategy for training end-to-end models. Website: montrealrobotics.ca/gigapixe… GitHub: github.com/montrealrobotics/… Great collaboration between MILA, NYU, Princeton and Torc Robotics:@rogg1111,@rdescc,@daphne_cor,@GrandhiAlaap,@_FelixHeide_,@EugeneVinitsky,@chrisjpal,@duckietown_coo

Excited to introduce #TruckDrive 🚛 at #CVPR2026: a new long-range driving dataset built specifically for long-range truck autonomy, where safe braking and anticipatory planning demand perception hundreds of meters ahead, far beyond existing robotaxi datasets. 📦 TruckDrive includes: 🔹 475K samples, with 165K densely annotated frames 🔹 Benchmarks for end-to-end driving, tracking, planning, depth estimation, and up to 1,000m for 2D detection and 400m for 3D detection 📏🎯 🛰️ A purpose-built long-range sensor suite: 🔸 7 long-range FMCW LiDARs (range + radial velocity) 🔸 3 high-res short-range LiDARs 🔸 11× 8MP surround cameras for short and long-range📷 🔸 10× 4D FMCW radars 📡 ⚠️ Key finding: current state-of-the-art models break down at long range 📉 with 31% to 99% drops on 3D perception tasks beyond 150m. TruckDrive exposes a long-range generalization gap that current architectures and training signals are not closing yet - a benchmark for the next generation of long-range highway autonomy research 🚚 🔗 Project and Data: torc-ai.github.io/TruckDrive… Fun work together with @torc_robotics led by Filippo Ghilotti, Edoardo Palladin, Samuel Brucker, Adam Sigal, and Mario Bijelic.

Are we done with object detection? What about tiny objects beyond 200 meters? 🔎 Telescope 🔭 addresses long-range perception by explicitly tackling extreme scale imbalance ⚖️ in images. It hinges on a learnable hyperbolic foveation transform from a low-resolution image, magnifying distant regions 🔍 while compressing nearby ones - effectively normalizing object scales with minimal computational overhead. Objects are detected in the transformed (Riemannian) space using a novel bounding box parameterization and are then mapped back to the original image. Project: light.princeton.edu/telescop…

ScenarioControl 🚗🛣️ - Scenario Generation from a single Dashcam Image 📸 or Text Prompt 💬!! Excited to introduce a new vision-language control mechanism for learned driving scenario generation. Given a single dashcam image or a scene prompt or an image, we generate a full scene layout 🧩, temporally consistent rollouts, including map 🗺️, agents 🚗, and ego video🛣️ ScenarioControl enables direct, fine-grained control over layout and traffic while preserving realism. It operates in a vectorized latent space with a new cross-global control mechanism to fuse vision-language inputs with scene structure while preserving realism. Interfaces seamlessly with generative video models! Project: light.princeton.edu/Scenario… Super fun project by Lili Gao, @Yanbo_Xu_ , William Koch, Samuele Ruffino, @Luke22R , Behdad Chalaki, Dmitriy Rivkin, Julian Ost, @rogg1111, Mario Bijelic.

Chop the gradients ✂️! We found that truncating decoder gradients in latent video diffusion to a fixed window allows us to finetune on videos with pixel-wise perceptual losses without running out of memory. Pixel losses have been essential for image generation and reconstruction, but until now, they haven't scaled to long-duration, high-resolution video diffusion due to recursive activation accumulation in causal decoders, leading to OOM during training 💥📉. Project: light.princeton.edu/chopgrad… Video diffusion models can do a lot more 🚀 when you can backprop the decoder! Post-process neural rendered scenes, super-resolve videos, harmonize lighting in controlled synthetic driving scenes, and inpaint videos — all in a single step ⚡ with a quick finetune from a standard diffusion model.

WorldFlow3D: Unbounded 3D World Generation 🌍 by Flow Through Hierarchical Distributions, without VAEs ! We reformulate 3D generation as flowing through sequentially finer 3D distributions, cutting training time by more than half ⏱️ compared to existing approaches! Vectorized map layouts provide full scene controllability 🗺️, and a novel flow-field alignment process enables causally coherent, spatially unbounded generation 🌍. This generative method generalizes across both real and synthetic data distributions! Project: light.princeton.edu/worldflo… Project led by @amogh7joshi and Julian Ost — will be super fun to build on this! 🔥

Excited to host ICCP 2026 in Princeton this summer!

Are we really done with autonomous driving 🚚? Remember the massive winter storm in the US last week❄️! We’re excited to share a large adverse weather driving dataset which includes small, distant road hazards, pushing perception beyond clear-weather and in-domain assumptions! light.princeton.edu/datasets… We collect with different imaging modalities, spanning LiDAR, RGB, gated imaging, stereo, polarization, and depth — collected across diverse weather, lighting, and range conditions, including rare adverse events like heavy rain (~5×/year) and dense fog (~12×/year in North America & Europe) that are typically underrepresented in standard driving benchmarks. What’s included: • Seeing Through Fog – labeled adverse weather dataset captured in over 10,000km of driving. • Gated2Depth / Gated2Gated – gated imaging for dense depth estimation. • Pixel-Accurate Depth Benchmark – ultra-high-resolution depth ground truth. • Long-Range Stereo (Gated Stereo) – large-scale sequential dataset with LiDAR and stereo (RGB, RCCB, gated). • Fogchamber Benchmark – long-range fog/rain depth benchmark. • Too Tiny To See – lost cargo benchmark with captures on snowy Lapland roads. • ScatterNeRF – scene reconstruction under atmospheric scattering. • Polarization Wavefront LiDAR – polarimetric LiDAR data. Exciting work coming out of a collaboration in the AI-SEE Project, @torc_robotics , @MercedesBenz, and @Princeton .

Physically-grounded video generations at #CES2026 without hallucinations! This week, we demo our end-to-end neural rendering developed at @torc_robotics, which allows us to simulate camera/lidar/radar data for edge cases, such as crash scenarios, road debris, or unforeseen crash sites, without hallucinations! We show our stack together with our friends from @ForetellixHQ  for end-to-end testing of driving models with reconstructed, generated, or conventional mesh assets. Join us at the Foretellix Booth at LVCC West Hall Booth 4767 from Tuesday to Friday in Las Vegas to see the demo!

Starting the new year without human labeling 🎉!! Multimodal lidar-camera data is a gold mine of dense 3D geometry hiding in plain sight. For supervised pretraining and validation at scale at @torc_robotics, we rely on fully automated pseudo-labeling pipelines. Exploiting geometric priors from temporally accumulated LiDAR maps and an iterative update rule enforces joint geometric–semantic consistency while detecting moving objects via inconsistencies. We achieve 3D semantic labels and 3D bounding boxes with human-like quality at 200m+ range required for highway driving. Paper: light.princeton.edu/unilips/ Exciting work with @torc_robotics with Filippo Ghilotti, Samuel Brucker, Nahku Saidy, Matteo Matteucci, Mario Bijelic.

Excited to share our #NeurIPS2025 work on learning motion hierarchies! We introduce a general hierarchical graph learning method that learns structured, interpretable motion directly from data, no prior structure or assumptions needed!!! Project and Paper: light.princeton.edu/HEIR Amazing work led by William Koch, @ChengZh1005, and @MotaLee5 ! See us in San Diego for #NeurIPS2025!