Lan Feng

PhD Candidate · VITA Lab, EPFL

I work on generative models and world models for physical AI. Physical AI is limited by the cost of the experience it learns from, and my research lowers that cost. RDM turns the four-step FLUX.2 [klein] into a one-step model that beats it on GenEval (0.826 vs 0.794). PTD, from my NVIDIA internship, distills the 33B audio-video model MiniMax-H3 to four steps with LoRA alone, cutting sampling time for a five-second clip from 343 s to 31 s. On the data side, TAROT chooses which data is worth training on.

Advised by Prof. Alexandre Alahi. Research intern at NVIDIA Research since June 2026. First or co-first author at ICLR 2026, ICML 2025 and ECCV 2024.

Open to research internships in 2027 Zurich, Switzerland

About

I am a PhD candidate at EPFL, Switzerland, working with Prof. Alexandre Alahi at the VITA Lab. Before EPFL I received my M.Sc. in Robotics, Systems and Control from ETH Zurich and B.E. in Navigation Engineering from Wuhan University.

I have built synthetic training data for physical AI since before my PhD. For my master's thesis at the AIT Lab, ETH Zurich, with Prof. Otmar Hilliges and NVIDIA, I trained human-to-robot handover policies on synthetic hand-object motions (SynH2R). Earlier, with Prof. Bolei Zhou at UCLA, I worked on generative simulation for reinforcement learning.

My current research makes generation itself cheap. RDM trains one-step generators by matching feature distributions under frozen encoders, and turns the four-step FLUX.2 [klein] into a one-step model that beats its teacher. PTD, from my internship at NVIDIA Research, distills the 33B audio-video model MiniMax-H3 into a four-step generator with LoRA alone; its videos are more diverse and more natural than those from DMD-style distillation, at matching motion quality. On the data side, TAROT (ICML 2025) selects training data with optimal transport, and RAP (ICLR 2026) shows that policies can be trained on cheap 3D rasterizations instead of photorealistic renders. I have also worked on minute-scale human animation (EverAnimate, NeurIPS 2026) and on world models that let a robot plan toward goals given in language (Grounded World Model).

Next, I am combining these into world models that are fast enough for closed-loop training and that generate the situations a policy fails on. I am looking for a research internship in 2027, and I am open to collaborations on these topics.

News

Oct 2026🚀 PTD is out: few-step video generation by parametric trajectory distillation, from my internship at NVIDIA Research. Four-step generation for the 33B audio-video MiniMax-H3, trained with LoRA alone. Project · arXiv.
Jul 2026🚀 RDM is out: one-step visual generation by distribution matching, turning the four-step FLUX.2 [klein] into a one-step model. Project · Demo.
Apr 2026🔜 Joining NVIDIA Research as a research intern for summer 2026, working on world models.
Jan 2026🎉 RAP accepted to ICLR 2026.
Oct 2025🚀 RAP: 3D Rasterization Augmented End-to-End Planning is on arXiv. It was the basis of our first-place entry in the 2025 Waymo Vision-based End-to-End Driving Challenge and ranked first on NAVSIM v1/v2 and Bench2Drive at release.
Jun 2025🏆 Won 1st place in the 2025 Waymo Open Dataset Vision-based End-to-End Driving Challenge.

Selected Publications

* equal contribution. Full list on Google Scholar.

Generative models

Parametric Trajectory Distillation for Few-Step Video Generation

Lan Feng, Peter Karkus, Maximilian Igl, Julius Berner, Yuxiao Chen, Shuhan Tan, Alexandre Alahi, Boris Ivanovic, Marco Pavone

arXiv 2026 · w/ NVIDIA Research

Four-step PTD on MiniMax-H331 s instead of 343 s per clip

Few-step distillation for large video diffusion and flow models: the student fits each denoising segment with a polynomial curve under a single least-squares loss and keeps the teacher's architecture. On the 33B audio-video MiniMax-H3, PTD trained with LoRA alone samples in four steps instead of 50 (31 s instead of 343 s for a five-second clip with sound on H100 GPUs). Against the DMD-based LightX2V Turbo it raises judged seed diversity from 1.55 to 2.81 at matching motion quality, and blinded raters give it a 63.4% preference share.

arXiv Project

Representation Distribution Matching for One-Step Visual Generation

Lan Feng, Wuyang Li, Éloi Zablocki, Matthieu Cord, Alexandre Alahi

arXiv 2026

A red sports car on a rainy neon street, generated by the four-step FLUX.2 [klein] teacher (left) and by the one-step RDM student (right)

Teacher, 4 stepsRDM, 1 step

A one-step image generator trained by matching feature distributions under frozen encoders, with no online teacher, adversary, trajectory, or JVP. Sets the one-step state of the art on ImageNet under SW_r14, a Sliced-Wasserstein distance over 14 frozen encoders, and post-trains the four-step FLUX.2 [klein] into a one-step model that beats it on GenEval (0.826 vs 0.794) and PickScore (22.76 vs 22.58) in 90 H200 GPU-hours.

arXiv Project Code ★ 80+ Weights Demo

EverAnimate: Minute-Scale Human Animation via Latent Flow Restoration

Wuyang Li, Yang Gao, Mariam Hassan, Lan Feng, Wentao Pan, Po-Chien Luan, Alexandre Alahi

NeurIPS 2026

A speaker at a podium after 90 seconds of animation: the Wan-Animate frame (left) shows background artifacts, the EverAnimate frame (right) stays clean

Wan-Animate at 90 sEverAnimate at 90 s

Minute-scale human animation with less drift: a persistent latent memory carries identity and motion across chunks, and restorative flow matching refines each chunk during sampling, with LoRA tuning only. At 90 seconds it reduces LPIPS by 32% and FID by 27% against prior long-animation methods.

arXiv Project Code

World models and data for physical AI

Grounded World Model: Latent Planning with Language Goals

Quanyi Li*, Lan Feng*, Haonan Zhang, Wuyang Li, Letian Wang, Alexandre Alahi, Harold Soh

arXiv 2026

“Pick up the tool used to clean shoes”

A latent world model that imagines the outcome of candidate actions and scores it against a language instruction in a shared vision-language space, so a robot can plan toward goals given in words. In simulation (WISER) it solves 87% of 288 tasks with unseen instructions, against 22% on average for ten fine-tuned VLA models, and on a real Franka arm it completes 55 of 60 pick-and-place sub-tasks.

arXiv

RAP: 3D Rasterization Augmented End-to-End Planning

Lan Feng, Yang Gao, Éloi Zablocki, Quanyi Li, Wuyang Li, Sichao Liu, Matthieu Cord, Alexandre Alahi

ICLR 2026 🏆 1st, Waymo 2025 Vision-based E2E Challenge

A camera frame from a driving log (left) and the matching lightweight 3D rasterization (right)

Camera3D rasterization

Data augmentation for end-to-end driving with lightweight 3D rasterization instead of photorealistic rendering: recovery maneuvers and cross-agent views, over 500k training samples, aligned to real images in feature space. Ranked first on NAVSIM v1/v2 and Bench2Drive at publication.

arXiv Project Code ★ 160+

TAROT: Targeted Data Selection via Optimal Transport

Lan Feng*, Fan Nie*, Yuejiang Liu, Alexandre Alahi

ICML 2025 🏆 nuScenes Prediction #1

A Cityscapes street scene with its segmentation map (left) and the nearest GTA5 frame by whitened feature distance, with its segmentation map (right)

Target, CityscapesNearest GTA5 frame

Selects the training subset whose optimal-transport distance to a target domain is smallest, measured with whitened feature distances. Beats prior selection methods on semantic segmentation, motion prediction, and LLM instruction tuning, and models trained on its subsets match or beat those trained on the full dataset.

arXiv Code

UniTraj: A Unified Framework for Scalable Vehicle Trajectory Prediction

Lan Feng*, Mohammadhossein Bahari*, Kaouther Messaoud Ben Amor, Éloi Zablocki, Matthieu Cord, Alexandre Alahi

ECCV 2024 🏆 nuScenes Prediction #1

Waymo, nuScenes, Argoverse 2 and MetaDrive data converted into one unified dataset, shown as top-down maps with trajectories

Unifies datasets, models, and metrics across Waymo, Argoverse, and nuScenes for vehicle trajectory prediction. Shows that models degrade sharply on unseen datasets, and that training on more, and more diverse, data gives a new state of the art on nuScenes.

arXiv Project Code ★ 480+

SynH2R: Synthesizing Hand-Object Motions for Learning Human-to-Robot Handovers

Sammy Christen*, Lan Feng*, Wei Yang, Yu-Wei Chao, Otmar Hilliges, Jie Song

ICRA 2024 · w/ ETH AIT Lab & NVIDIA

Synthetic human hands handing over an object in simulation (left) and a real robot arm taking a banana from a person (right)

Synthetic human motionReal robot

Trains human-to-robot handover policies on synthetic hand-object motions instead of motion-capture data. A handover-aware grasp synthesis method generates training and test data with 100x more objects than prior work, and the policy trained purely on synthetic data is competitive with methods that use real human motion, in simulation and on a real robot.

arXiv Project

ScenarioNet: Open-Source Platform for Large-Scale Traffic Scenario Simulation and Modeling

Quanyi Li*, Zhenghao Peng*, Lan Feng*, Zhizheng Liu, Chenda Duan, Wenjie Mo, Bolei Zhou

NeurIPS 2023 Datasets and Benchmarks Track

A real driving scene with annotated vehicles (left) and the same scenario replayed in the MetaDrive simulator (right)

Real-world logReplayed in simulation

An open-source platform that converts real-world driving logs from Waymo, nuScenes, Lyft L5 and nuPlan into one scenario format and replays them as digital twins in the MetaDrive simulator, from bird’s-eye-view layouts to 3D rendering. Supports large-scale scenario generation, imitation learning and reinforcement learning.

arXiv Project Code ★ 290+

Other work

Weak-for-Strong: Training Weak Meta-Agent to Harness Strong Executors

Fan Nie, Lan Feng, Haotian Ye, Weixin Liang, Pan Lu, Huaxiu Yao, Alexandre Alahi, James Zou

COLM 2025 · w/ Stanford

arXiv

OmniTraj: Pre-Training on Heterogeneous Data for Adaptive and Zero-Shot Human Trajectory Prediction

Yang Gao, Po-Chien Luan, Kaouther Messaoud, Lan Feng, Alexandre Alahi

Transportation Research Part C, 2026

Paper arXiv

TrafficGen: Learning to Generate Diverse and Realistic Traffic Scenarios

Lan Feng*, Quanyi Li*, Zhenghao Peng*, Shuhan Tan, Bolei Zhou

ICRA 2023 · w/ UCLA

arXiv Project Code ★ 190+

MetaDrive: Composing Diverse Driving Scenarios for Generalizable Reinforcement Learning

Quanyi Li*, Zhenghao Peng*, Lan Feng, Qihang Zhang, Zhenghai Xue, Bolei Zhou

IEEE TPAMI, 2022

arXiv Code ★ 1.2k+

Human-AI Shared Control via Policy Dissection

Quanyi Li, Zhenghao Peng, Haibin Wu, Lan Feng, Bolei Zhou

NeurIPS 2022

arXiv

Awards and Honors

2025

🏆 1st Place, Waymo Open Dataset Vision-based End-to-End Driving Challenge

$10,000 prize. Entry UniPlan, with Alexandre Alahi (CVPR 2025 workshop results).
2021

🏆 1st Place, Waymo Open Dataset Interaction Prediction Challenge

$15,000 prize.
2021

🥉 3rd Place, Waymo Open Dataset Motion Prediction Challenge

$2,000 prize.
2019

🎓 China National Scholarship

Experience & Education

Experience

Jun 2026 — Mar 2027

Research Intern · NVIDIA Research

ASPIRE group (Autonomous Systems and Physical AI Research), Zurich. Few-step video generation for world models (PTD). Full-time through Sep 2026, part-time since. Worked with Peter Karkus, Maximilian Igl, Yuxiao Chen, Boris Ivanovic and Marco Pavone.
Nov 2023 — now

Graduate Researcher · VITA Lab, EPFL

Advised by Prof. Alexandre Alahi.
Feb 2022 — Sep 2023

Research Assistant · AIT Lab, ETH Zurich

RL for dexterous grasping; master's thesis on human-to-robot handovers with NVIDIA (SynH2R). Advised by Prof. Otmar Hilliges.
Nov 2020 — Aug 2021

Research Assistant · Zhou Lab, UCLA

Generative simulation for reinforcement learning (TrafficGen, ScenarioNet). Advised by Prof. Bolei Zhou.
Jan 2020 — Aug 2020

Trainee Researcher · SenseTime

Learning-based traffic simulation.

Education

May 2024 — now

PhD in Robotics, Control and Intelligent Systems · EPFL

Thesis on data selection & curation for foundation models. Advised by Prof. Alexandre Alahi.
Sep 2021 — Oct 2023

M.Sc. in Robotics, Systems and Control · ETH Zurich

GPA 5.67/6.0. Master thesis at AIT Lab w/ Prof. Otmar Hilliges and NVIDIA (SynH2R).
Sep 2016 — May 2020

B.E. in Navigation Engineering · Wuhan University

GPA 3.9/4.0.

Contact

The best way to reach me is by email at lan.feng@epfl.ch. I am based in Zurich, Switzerland. I also use X/Twitter and LinkedIn.

I am especially interested in collaborating on: (i) one- and few-step image and video generation, (ii) world models that generate training experience for physical AI, and (iii) choosing which data to generate and train on.