CoRL 2025

CoRL 2025

Conference on Robot Learning 2025 — Seoul, Sept 27–30, 2025.

263 accepted papers (42 orals + 221 posters). VLA/manipulation was the dominant thread; NVIDIA launched Isaac GR00T N1.6 + the Newton physics engine at the venue.

Surveys hosted in this wiki

• VLA & Manipulation Survey (CoRL 2025) — categorized summary of CoRL 2025's VLA/manipulation papers plus trends and a CoRL 2025 → ICLR 2026 lineage map.

Awards

Award	Paper
Best Paper	Fabrica — dual-arm general multi-part assembly
Best Paper	UniFP — unified position+force policy for legged loco-manipulation
Best Student Paper	Visual Imitation → Humanoid — contextual humanoid control from internet video
Best Paper Finalist	DexUMI, DSRL, LocoFormer, The Sound of Simulation, π0.5

Quick paper index (by category)

Flagship baseline

• π0.5 (Oral) — Physical Intelligence's hierarchical VLA with co-training; sets up the entire 2025→2026 π series. See π series evolution.

VLA architecture

• DexVLA — plug-in ~1B diffusion action expert atop a VLM
• TA-VLA — single torque-history token in the decoder for contact-rich tasks
• Streaming Flow Policy (Oral) — action chunk = point on a longer flow

Training & inference recipes

• ECoT-Lite — which parts of embodied chain-of-thought actually matter
• RoboMonkey — best-of-N VLA sampling with a learned verifier

Dexterous / humanoid / whole-body

• DexUMI (Finalist) — wearable "universal manipulation interface"
• ClutterDexGrasp (Oral) — zero-shot sim-to-real closed-loop dex grasping in clutter
• UniFP (Best Paper) — unified position+force legged loco-manipulation
• Visual Imitation → Humanoid (Best Student Paper) — everyday video → humanoid skills
• Fabrica (Best Paper) — dual-arm multi-part assembly

Cross-embodiment from human video

• X-Sim (Oral) — real-to-sim-to-real via object motion

Diffusion / flow policies + RL

• DSRL (Oral, Finalist) — RL in the initial-noise latent space of a frozen diffusion policy

World models for policy

• DreamGen — policy training inside a video world model (NVIDIA)

Data & benchmarks

• ManipBench — first VLM benchmark targeting low-level manipulation reasoning

Trends (6)

1. Hierarchical VLAs beat monolithic VLAs for open-world generalization (π0.5, OneTwoVLA, Long-VLA).
2. Test-time scaling + trajectory streaming are the cheap latency/quality levers (RoboMonkey, Streaming Flow Policy, DemoSpeedup, SAIL).
3. Human video is the default cross-embodiment data source (DexUMI, Visual Imitation → Humanoid, UniSkill, ImMimic, X-Sim).
4. Diffusion / flow policies get RL-ified without log-probs (DSRL, DiWA) — prefigures ICLR 2026's RECAP, RL Tokens, VLA-RFT, SimpleVLA-RL.
5. Contact / force is finally modeled inside VLAs (TA-VLA, DexSkin, UniFP, KineSoft, Tactile Beyond Pixels).
6. World models move inside (not adjacent to) training pipelines (DreamGen, LaDi-WM, FLARE, ParticleFormer) — with NVIDIA GR00T N1.6 + Newton as the ecosystem signal.

← Back to CoRL · Home