IROS 2025 - Heungwoo/research GitHub Wiki

IROS 2025

IEEE/RSJ International Conference on Intelligent Robots and Systems 2025 โ€” Hangzhou, China, October 19โ€“25, 2025.

Scale: 4,306 submissions โ†’ 1,991 main-track accepted (46.2%) + 777 RA-L / T-RO / T-ASE / T-MECH journal transfers. Record submission year. Five parallel "Award Finalists" sessions across ~45 papers.

Awards relevant to VLA / manipulation:

  • Best Conference Paper โ€” biosyncretic-robot electrodes (Shenyang Institute of Automation) โ€” not VLA-direct
  • Best Student Paper โ€” Neural MP (Dalal et al., CMU) โ€” neural motion planner using learned priors
  • Best Paper Finalists include VLA/manipulation papers:
    • AgiBot World Colosseo (OpenDriveLab + AgiBot) โ€” 1M+ trajectories across 100 robots
    • Interactive Legged-Manipulator Navigation
    • Unified Whole-Body Policy

Surveys hosted in this wiki

  • VLA & Manipulation Survey (IROS 2025) โ€” 18+ IROS 2025 VLA/manipulation papers in 7 categories. Flagship: AgiBot World Colosseo (1M+ trajectories, Best Paper Finalist) + Neural MP (Best Student Paper). Includes a CoRL 2025 โ†’ IROS 2025 โ†’ NeurIPS 2025 / ICLR 2026 lineage map.

Quick paper index (by category)

VLA architecture (IROS-specific spins)

  • RoboNurse-VLA (CUHK MSMRC) โ€” SAM 2 + Llama 2 for surgical scrub-nurse; voice-driven, real-time
  • CapsDT (CUHK Ren Lab) โ€” Diffusion-Transformer VLA for magnetically-actuated capsule endoscopy in the stomach
  • TinyVLA (Midea + ECNU) โ€” Compact VLA family (40Mโ€“1.4B-param backbones); TinyVLA-H beats OpenVLA by +25.7% success at 5.5ร— fewer params and ~20ร— lower inference latency
  • PD-VLA (HKUST-GZ + PolyU) โ€” Parallel fixed-point iteration reformulation of AR VLA decoding; 2.52ร— faster, training-free
  • ReBot (UNC + UT Austin) โ€” Real-to-sim-to-real video synthesis for VLA adaptation; +20% Franka real for OpenVLA
  • Refined Policy Distillation โ€” Distills a VLA into a compact RL expert

Real-robot deployment / systems

  • HACTS โ€” Bilateral teleoperation hardware
  • Bunny-VisionPro โ€” Vision Pro bimanual teleop
  • MoMa-Teleop โ€” Zero-added-cost whole-body mobile-manip teleop
  • iDP3 Humanoid โ€” 25-DoF cart-mounted humanoid + 3D diffusion policy

Dexterous manipulation

  • VTAO-BiManip โ€” Visuo-Tactile-Action pretraining for bimanual
  • LDexMM โ€” Language-conditioned dex multi-task
  • RoboDexVLM โ€” Dex manipulation task planning
  • DexDiffuser โ€” Diffusion for dex grasping
  • FoundationGrasp โ€” Foundation model for grasping

Mobile manipulation

  • MORE โ€” Mobile manipulation benchmark
  • MoMa-Teleop โ€” also real-robot deployment

Humanoid / whole-body

Data / benchmarks

  • AgiBot World Colosseo (Best Paper Finalist, Genie Operator-1) โ€” 1M+ trajectories, 217 tasks, 100 robots, 5 domains
  • Neural MP (Best Student Paper) โ€” Neural motion planner, CMU

Teleoperation & imitation-data interfaces

Relevant workshops

  • Generative AI for Robotics
  • Tactile Manipulation (hosts Tactile-VLA best paper)
  • AI Meets Autonomy
  • RoDGE (Robotic Dexterous Grasping and Execution)
  • AI Robot for Science
  • Fine Manipulation
  • Embodied AI for Science

Trends โ€” what's distinctive about IROS 2025

  1. VLAs as vertical-domain products โ€” RoboNurse-VLA (surgical), CapsDT (capsule endoscopy), RoboDexVLM (manipulator planning). IROS is where VLA meets real application with hardware.
  2. Real-robot systems integration, not architecture theory. Every paper has a hardware demo. HACTS, Bunny-VisionPro, MoMa-Teleop, iDP3-humanoid all feature hardware novelty as much as algorithm.
  3. VLA acceleration, compression, distillation. PD-VLA (2.52ร— faster), TinyVLA (<1B params), ReBot (real-to-sim-to-real), Refined Policy Distillation โ€” the "how do we actually ship this" papers.
  4. Industrial-scale data platforms. AgiBot World Colosseo's 1M+ trajectories dwarfs every prior manipulation dataset.
  5. Tactile is a whole research track at IROS. VTAO-BiManip, Tactile-VLA (workshop), Adaptive Visuo-Tactile Fusion, RoTipBot. IROS is where tactile research concentrates.

Cautionary notes

  • Harmonic Mobile Manipulation is IROS 2024, not 2025 โ€” commonly miscited.
  • IROS papers often have arXiv preprints from late 2024 / early 2025 (RA-L pipeline) and are presented at IROS in October. Chronological lineage claims should account for this.
  • IROS has ~40 category-specific awards, not a single Outstanding Paper list โ€” "Best Paper" contests are across multiple tracks.

โ† Back to IROS ยท Home