Chenggang Liu

IROS 2026

IROS 2026 at am Glance   export

Papers Distribution (count, % of 1,929 papers)

  • Based on themes
    Theme Papers Share
    Learning & AI (RL, imitation/demonstration, deep learning) 800 41.5%
    Navigation, Planning & Control 511 26.5%
    Locomotion & Platforms (legged, aerial, marine, field, soft) 385 20.0%
    Manipulation & Grasping 299 15.5%
    Perception & Vision 277 14.4%
    SLAM, Localization & Mapping 229 11.9%
    Multi-Robot & Human-Robot Interaction 227 11.8%
    Sensing & Tactile 154 8.0%
    Safety 128 6.6%
    Mechanism & Design 94 4.9%
    Medical & Assistive 88 4.6%
  • Based on keywords (count, % of 1,929 papers)
    Keyword Count %
    Reinforcement Learning 198 10.3%
    Motion and Path Planning 152 7.9%
    Deep Learning for Visual Perception 144 7.5%
    Imitation Learning 121 6.3%
    Deep Learning in Grasping and Manipulation 117 6.1%
    Vision-Based Navigation 117 6.1%
    Machine Learning for Robot Control 101 5.2%
    Multi-Robot Systems 95 4.9%
    Deep Learning Methods 91 4.7%
    AI-Enabled Robotics 88 4.6%
    Mapping 88 4.6%
    Sensor Fusion 81 4.2%
    Legged Robots 80 4.1%
    Learning from Demonstration 80 4.1%
    SLAM 76 3.9%
    AI-Based Methods 76 3.9%
    Optimization and Optimal Control 75 3.9%
    Localization 75 3.9%
    Autonomous Vehicle Navigation 74 3.8%
    Force and Tactile Sensing 74 3.8%
  • Based on applications

    About 57% of papers are general methods that state no specific application.

    Application domain Papers Share
    Self-driving / Autonomous vehicles 182 9.4%
    Humanoid / Legged locomotion 152 7.9%
    Aerial / Drones / UAV 122 6.3%
    Human-Robot Interaction / Collaboration 122 6.3%
    Medical / Surgical / Healthcare 92 4.8%
    Marine / Underwater 54 2.8%
    Assistive / Rehab / Wearable / Prosthetics 39 2.0%
    Manufacturing / Industrial 35 1.8%
    Home / Service / Domestic 32 1.7%
    Search & Rescue / Disaster / Field 28 1.5%
    Agriculture / Forestry 28 1.5%
    Warehouse / Logistics 16 0.8%
    Construction / Inspection 12 0.6%
    Space / Planetary 11 0.6%
  • Based on institutions (all papers)
    Institution Papers Share
    Tsinghua University 91 4.7%
    Carnegie Mellon University 77 4.0%
    Zhejiang University 60 3.1%
    HKUST 54 2.8%
    National University of Singapore 50 2.6%
    Shanghai Jiao Tong University 50 2.6%
    Chinese University of Hong Kong 45 2.3%
    Purdue University 45 2.3%
    MIT 43 2.2%
    University of Tokyo 43 2.2%
    Georgia Tech 36 1.9%
    KAIST 36 1.9%
    Northeastern University 36 1.9%
    Technical University of Munich 36 1.9%
    Harbin Institute of Technology 33 1.7%
    Nanyang Technological University 32 1.7%
    Beijing Institute of Technology 32 1.7%
    University of Michigan 31 1.6%
    SUSTech 31 1.6%
    University of Southern California 28 1.5%
    Beihang University 27 1.4%
    Peking University 27 1.4%
    ETH Zurich 26 1.3%
    Seoul National University 25 1.3%
    Tongji University 24 1.2%
    Institute of Automation, CAS 24 1.2%
  • Based on country / region
    Country Papers Share
    China 571 29.6%
    USA 528 27.4%
    Japan 119 6.2%
    Germany 105 5.4%
    Korea 87 4.5%
    Singapore 80 4.1%
    UK 71 3.7%
    Italy 42 2.2%
    Switzerland 41 2.1%
    Canada 31 1.6%
    France 27 1.4%
    Netherlands 14 0.7%
    Other / Unresolved 456 23.6%

Self-driving Papers Distribution (count, % of 182 papers)

Note: the method table below is over the 62 self-driving ML-planner papers (a subset of the 182), not all 182.

  • Based on sub-area
    Sub-area Papers Share
    Navigation (high-level: AV navigation, ITS) 104 57.1%
    Perception (detection, segmentation, scene understanding) 67 36.8%
    Learning methods (RL / DL / representation / imitation) 55 30.2%
    Localization, Mapping & SLAM 41 22.5%
    Planning & Decision-making 36 19.8%
    Prediction & Forecasting 29 15.9%
    Safety 19 10.4%
    Simulation & Benchmarking 14 7.7%
    Multi-agent / V2X 7 3.8%
  • Based on method
    ML method family Papers Share of 62
    Reinforcement Learning (RL) 16 25.8%
    VLM / LLM / foundation model 15 24.2%
    Diffusion / score-based 14 22.6%
    Transformer / attention 14 22.6%
    Imitation / Behavior Cloning 6 9.7%
    MPC + learning / hybrid control 6 9.7%
    World model / model-based 5 8.1%
    Autoregressive / sequence / tokens 4 6.5%
    Vector quantization / codebook 2 3.2%
    Flow matching / rectified flow 2 3.2%
    Graph neural network 1 1.6%
    GAN / VAE / generative (other) 1 1.6%
    Inverse RL / IOC / cost learning 1 1.6%
    (no specific method matched) 11 17.7%

Exhibitor distribution (IROS 2026 exhibit hall)

Field Companies Share
Humanoid / general-purpose embodied robots 24 13.6%
AI / data / simulation / software (physical AI) 24 13.6%
Manipulators / industrial / cobots 18 10.2%
Education / associations / publishers 15 8.5%
Dexterous hands / grippers / end-effectors 14 7.9%
Tactile / force / haptic sensing 14 7.9%
Actuators / motors / drives / components 12 6.8%
Big tech / industry / government labs 11 6.2%
LiDAR / depth / vision sensors 9 5.1%
Motion capture / tracking 8 4.5%
Quadruped / legged / aerial / field 8 4.5%
Medical / assistive / BCI / prosthetics 6 3.4%
Other / uncategorized 6 3.4%
Autonomous vehicles / driving 4 2.3%
Semiconductors / edge / power / hardware 4 2.3%

Self-driving 'Research' North Star

1. End-to-end

  • [4895] NPPC: Neural Parametric Planning Cost for End-To-End Autonomous Driving   P0 export R0
    • Summary
      20261015_074807.png

      Problems:

      • E2E planners trained by behavior cloning (imitating expert trajectories) suffer from biases and compounding errors in closed loop.
      • Inverse Optimal Control (IOC) can help by learning the expert's planning cost, but
        • prior IOC uses hand-crafted cost terms, may not fully capture nuanced human decisions.
        • spatially discrete cost-maps (need rule-based post-processing), they may not accurately capture high-risk regions in the continuous driving space.
        • High-dim neural scorers (limited to ranking predefined trajectory candidates), limited to scoring a set of predefined trajectory proposals during inference, restricting flexibility in the final trajectory.

      How:

      • predicts a spatio-temporal cost-map as a weighted sum of radial basis functions (RBFs) conditioned on the scene

      Why interesting:

      • parametric learnable cost
      • combine rule-based cost and learned cost
    • Method
      20261004_124245.png
      • NPPC predicts a spatio-temporal cost-map as a weighted sum of radial basis functions (RBFs) conditioned on the scene (low-dimensional, spatially continuous, learnable cost).
      • Because it's low-dim, it plugs directly into downstream planners as an optimization objective with no post-processing, enabling maneuvers beyond predefined candidates.
      • IOC
        • cost function: \(g(\cdot) = \sum_{k=1}^{K}w_{k}\phi(\cdot)\)
        • RBF: \(\phi(x, \mu, \Sigma) = \frac{1}{1 + (x - \mu)^\top \Sigma^{-1}(x-\mu)}\), where \(\mu\) is xy-coordinate.
        • loss: max-margin loss which is the same as Maximum Margin Planning.
        • Sampler: engineered sampler, which could be limit factor.
        • Refer to End-to-end Interpretable Neural Motion Planner https://arxiv.org/pdf/2101.06679
      • Planner: sampling, ranking, and optimization for refinement. The refinement step is a major difference from previous work as shown in Fig. 1.
    • Overall

      max-margin IOC with an learnable/interpretable cost using RBF and optimization for final refinement.

2. Multimodal trajectory generation

  • [706] G2DP: Diffusion Planning with Spatio-Temporal Grid Guidance   P0 export
    • Summary
      20261016_090544.png
      • Problem:
        • Diffusion planners model diverse behaviors well but needs explicit guidance for safety/route adherence.
        • Existing guidance using sparse entity-centric geometric queries or post refinement is not good at situational awareness and being fragile in interactive scenes.
      • How
        • builds a differentiable spatio-temporal cost volume and injects dense gradients into the denoising loop
      • Why interesting:
        • representation/guidance idea rather than a new generator
    • Method
      20261004_143552.png

      G2DP (Grid-Guided Diffusion Planning) enforces dense environmental constraints via inference-time guidance. It builds a differentiable spatio-temporal cost volume by fusing probabilistic future occupancy with a route-progress map, formulates it as a continuous safety energy functional, and injects dense gradients into the denoising loop to steer trajectories toward collision-free, progress-optimal regions.

      • U-net to predict Occ map
      • Fuse with route-aligned progress (encourage the ego to move forward and stay in the lane): \(\psi_{t}(u, v) = \gamma O_{t}(u, v) + (1-\gamma) R(u, v)\)
      • To balance safety coverage with robustness, we propose a Top-K footprint aggregation strategy.
      • Gradient-Based Denoising Steering.
    • Result:

      SOTA on nuPlan (+7.2 reactive score over the strongest IL baseline); top zero-shot transfer to interPlan and DeepScenario (collision avoidance +10.15 over unguided on interPlan). Code available (HangYuu/G2DP). TUM / Mercedes-Benz / KIT. IROS 2026.

      • Overall: well-rounded closed-loop result with strong zero-shot transfer (evidence of generalization); contribution is a representation/guidance idea rather than a new generator, so impact hinges on how much the dense grid beats cheaper guidance.
  • [398] Vector Quantization for Trajectory Prediction   P0 export
    • Summary
      20261004_150211.png
      • Problem: motion-forecasting decoders stayed simplistic.
        • Regression decoders (Winner-Takes-All) suffer training instability and mode collapse;
        • heuristic classification over hand-crafted clusters/goals gives good mode coverage but poor absolute accuracy;
        • autoregressive decoders are slow (token-by-token). Planners mostly act on the single most-likely mode, so that mode's quality matters most.
      • How A novel trajectory decoding framework that transforms motion forecasting into a classification problem by learning discrete latent trajectory priors via Vector Quantization (VQ).
      • Why interesting
        • interpretable decision encoder (Codebooks).
    • Method
      20261004_153100.png

      How to train vector quantizer and trajectory decoder

      20261004_153113.png

      How to train scene encoder

      VQTraj frames trajectory decoding as classification over discrete latent trajectory priors learned via Vector Quantization. An autoencoder with finite-scalar quantization (FSQ) learns a codebook where one code encodes a full trajectory (interpretable, data-driven priors instead of heuristics). Pretrain the autoencoder to reconstruct futures, then reuse it as a drop-in decoder: adapted models just predict a distribution over latent priors.

      • Autoencoder to c, then finite-scalar quantization (FSQ) to get discrete codebook,
    • Result

      Upgrades two SOTA models on Argoverse 2, improving both the top-1 prediction (minADE1, minFDE1) and ranked multimodal predictions (bminFDE6); strong zero-shot cross-dataset transfer to Waymo Open Motion. Waabi / Univ. Toronto. IROS 2026.

5. A safety layer

  • [82] Learning a Unified Risk Map for Autonomous Driving   R0 export
    • Problem
      20261014_115524.png
      • Occlusion-aware prediction is hard under unobserved-region uncertainty.
      • Prior methods either
        1. overestimate risk (reachable-set, overly conservative) or
        2. predict poorly under high occlusion (trajectory/occupancy prediction).
      • Occluded-interaction scenarios are also scarce in data.
      • Key insight:
        • model a dense, planner-friendly unified spatio-temporal risk field that implicitly marginalizes over the uncertainty
        • solve the data-scarcity problem with adversarial diffusion-based scenario generation.
      • Why interesting
        • prediction-planning interface (model risk vs trajectory)
        • interactive data generation
        • predict risk
    • Method: four interconnected components (Fig. 2).
      20261014_104302.png
      • Occlusion risk modeling (the risk field, 0.5 m grid \(\Omega\)):
        • Flow risk: spatial density of predicted multimodal trajectories (where traffic is likely), distance-decayed by \(e^{-\lambda D}\); stationary agents filtered by a speed threshold \(v_{min}\).
        • Collision risk: built from spatio-temporal overlap events where the ego trajectory comes within \(\delta\) of a predicted trajectory.
        • Fused linearly \(R_{total} = \alpha R_{flow} + \beta R_{collision}\), then Gaussian-filtered and normalized.
      • Occlusion interaction data generation (diffusion):
        • samples hidden-agent initial states (position/speed) in occluded lane segments from map topology + ego FOV;
        • a pretrained DDPM (transformer scene encoder + denoiser) predicts controls (accel \(\dot{v}\), yaw rate \(\dot{\psi}\)) rolled out via a bicycle model.
        • A guidance function \(F(u^p) = \lambda_1 \min_t d_{inter} + \lambda_2 \min_t d_{road}\) steers generation toward adversarial-but-realistic (maximize interaction risk vs. SDF road adherence).
      • Risk prediction network (lightweight, real-time ~6.67 ms / 150 FPS):
        • ray-traced FOV encoding \(f_{fov}\) + vectorized map \(f_{map}\) fused by cross-attention;
        • lane anchors as queries (QCNet-style) in a transformer decoder output per-anchor, multi-step risk scores (trained with MSE on modeled risk).
      • Driving strategy: QP planner with composite cost \(C_{total} = w_1 C_{smooth} + w_2 C_{reach} + w_3 C_{risk} + w_4 C_{collision}\), where \(C_{risk} = \sum_i R_i v_i^2\) discourages speed in risky regions (slow down before a blind spot).

Workshops

Other Resource   export