Compiled April 2026 (in retrospect). Focus: VLA architecture, training & inference recipes, dexterous / humanoid manipulation, cross-embodiment, diffusion & flow policies, world models, benchmarks.

CoRL 2025 (Seoul, Sept 27–30) accepted 221 papers out of 263 submissions (84.03% acceptance), of which 42 were orals. From the VLA/manipulation perspective, six trends stand out:

1. Hierarchical VLAs beat monolithic ones for open-world generalization — exemplified by π0.5 (CoRL 2025 Oral + Best Paper Award Finalist), which sets up the entire π series that the field now benchmarks against.
2. Test-time scaling + trajectory streaming — RoboMonkey (best-of-N with verifier), Streaming Flow Policy, DemoSpeedup, SAIL.
3. Human video as 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 (noise-latent RL), DiWA — the seed for ICLR 2026's explosion of RL-for-flow-matching.
5. Contact / force becomes first-class — TA-VLA (single torque token in decoder), DexSkin, Tactile Beyond Pixels, UniFP (Best Paper).
6. World models move inside training — DreamGen trains policies inside video-world-model rollouts; NVIDIA's GR00T N1.6 + Newton + Cosmos are the ecosystem analog.

flowchart LR
  V[Vision] --> B[VLM Backbone]
  L[Task instruction] --> B
  B --> H[High-Level<br/>Subtask Predictor]
  B --> AE[Flow-Matching<br/>Action Expert]
  H --> AE
  AE --> A[Continuous actions]
  WD[Web data<br/>+ detection<br/>+ subtask text] -. co-training .-> B

π0.5 (Physical Intelligence, CoRL 2025 Oral + Best Paper Award Finalist) is the flagship CoRL 2025 paper for this wiki's purposes. It adds a high-level subtask predictor on top of π0's flow-matching action expert, co-trains on heterogeneous data (multi-robot teleop + web VQA + detection + subtask semantic prediction), and is the first VLA to demonstrate long-horizon manipulation (cleaning kitchens, bathrooms) in entirely unseen homes.

Every subsequent π release builds on π0.5: π0.6 (Nov 2025, Gemma3-4B + Knowledge Insulation), π*0.6 + RECAP (Nov 2025, RL-from-experience), π0.7 (Apr 2026, MEM history + subgoal-image world model + metadata prompting). See π series evolution.

1. VLA Architecture
2. Training & Inference Recipes
3. Dexterous / Humanoid / Whole-Body Manipulation
4. Cross-Embodiment from Human Video
5. Diffusion / Flow Policies + RL
6. World Models for Policy Learning
7. Data & Benchmarks
8. Trends & Key Findings
9. CoRL 2025 → ICLR 2026 Lineage
10. Reading List

Also worth knowing: FLOWER (efficient VLA for consumer hardware), 3DS-VLA (3D spatial features inside a VLA), Long-VLA (long-horizon planning tokens), OneTwoVLA (reasoning + acting dual mode).

Human video is the default data source at CoRL 2025 for scaling across embodiments:

• X-Sim (Oral) — real-to-sim-to-real via object motion as the embodiment-agnostic supervisory signal (Cornell).
• UniSkill — embodiment-invariant skill embedding learned from human video.
• ImMimic (Oral) — interpolation between human and robot spaces for cross-domain imitation.
• DexUMI — already listed; wearable human-hand interface is a human-video analog for dexterous data.

The 2025 insight: diffusion policies have latent structure you can do RL on without ever computing action log-probs.

flowchart LR
  N[Initial noise z0<br/>RL action space] --> D[Frozen diffusion policy]
  D --> A[Action chunk]
  A --> R[Real-world reward]
  R -- update z0 policy --> N

• DSRL (Oral, Finalist) — the headline result: initial-noise latent encodes trajectory info; RL in that latent space steers a frozen diffusion policy cheaply and stably.
• DiWA — adapts a pretrained diffusion policy using rollouts in a learned world model.
• Belief-Conditioned One-Step Diffusion (Oral) — trades denoising steps for belief conditioning to hit real-time budgets.
• LatentToM (Oral) — decentralized diffusion policy architecture for cooperative multi-robot.
• D-CODA — diffusion-based data augmentation targeting bimanual coordination scarcity.

This whole direction is the seed of ICLR 2026's explosion of RL-for-flow-matching: RECAP, RL Tokens, SimpleVLA-RL, VLA-RFT, Stage-Aware RL.

Ecosystem signal: NVIDIA launched Isaac GR00T N1.6 + Newton physics engine + Cosmos simulation/generation tools at CoRL 2025. These thread into ICLR 2026 via Ctrl-World, Cosmos Policy, WorldGym.

• ManipBench — 12,617 MCQs across 33 VLMs / 10 families, first benchmark squarely on low-level manipulation reasoning (object–object interaction, deformables). Foreshadows ICLR 2026's VLM4VLA finding that VLM scores ≠ manipulation skill.
• BEHAVIOR Robot Suite — reference platform + benchmark for whole-body household tasks.
• Tactile Beyond Pixels (Meta FAIR + GT) — pretrains multimodal touch representations from ~1M contact-rich interactions across four tactile modalities.

1. Hierarchy wins in the open world — π0.5's subtask-head + action-expert split is the template every serious generalist VLA now follows.
2. Test-time scaling arrives for VLAs — RoboMonkey-style best-of-N is cheap and composable.
3. Human video = cross-embodiment data — wearable interfaces, retargeting, object-motion supervision all converge.
4. RL for diffusion policies, sans log-probs — DSRL is the seed idea; every ICLR 2026 RL-for-VLA paper can trace an ancestor here.
5. Contact/force is now first-class — single-token torque conditioning, full-hand tactile, unified force+position.
6. World models inside training — DreamGen and NVIDIA's stack are the production analog.

Direct descendant threads (links go to ICLR 2026 pages):

Papers I could not confirm as CoRL 2025 (intentionally absent): Octo (ICRA 2024), OpenVLA (CoRL 2024), RDT-1B (ICLR 2025), HPT (NeurIPS 2024). Helix (Figure AI) and Gemini Robotics (DeepMind) were industry demos around the venue, not accepted CoRL 2025 papers.

• CoRL 2025 official site
• CoRL 2025 awards
• Stanford AI Lab CoRL 2025 summary
• Curated 221-paper aggregator
• Paper Copilot — CoRL 2025 stats
• Paper Copilot — CoRL 2025 paper list
• Lumeny attendee trends summary
• NVIDIA at CoRL 2025 (GR00T N1.6 + Newton + Cosmos)
• The Robot Report — Newton / GR00T N1.6 launch

Per-paper links appear on each paper's page.

← Back to CoRL-2025 · CoRL · Home