ZJU's VLA-Δ Sets 98.4% LIBERO Record, SJTU/HKU's OneWM Doubles π0.5 on RoboCasa365, and CoDA Lifts Perturbed Robot Success from 1% to 76%

Core Background & Industry Pain Points

While generalist Vision-Language-Action (VLA) models (OpenVLA, π0, π0.5, GR00T) achieve high scores on in-distribution simulation benchmarks, they frequently fail under out-of-distribution (OOD) visual, spatial, and linguistic perturbations. Three structural limitations drive this fragility: (1) direct observation-to-action mapping without explicit supervision of intermediate physical state transitions; (2) isolated training of video world models and action policies without exploiting their reciprocal causal synergy; and (3) open-loop generative simulation engines that synthesize random environments rather than targeting specific policy failure modes with dense value supervision.

Architectural Highlights & Core Mechanisms

Three new open-source embodied AI frameworks—Zhejiang University's VLA-Δ (arXiv:2610.21306, 54 upvotes), SJTU and HKU's OneWM (arXiv:2610.21095, 40 upvotes), and UIUC/Amazon's CoDA (arXiv:2610.22362)—address these limitations across pretraining, causal architecture, and closed-loop data generation:

• ZJU's VLA-Δ: Zero-Inference-Overhead Physical State Transition Supervision: Pretrains VLA backbones using explicit task-relevant semantic differences (Future-State Difference Text) and visual motion changes between current and future states as learning targets. At deployment, VLA-Δ generates continuous actions directly from current observations with zero additional inference latency.
• SJTU & HKU's OneWM: Unified Causal World and Action Modeling: Treats actions simultaneously as conditioning inputs for forward video dynamics prediction and as targets for policy learning within a shared causal video-action architecture, demonstrating mutual reinforcement between world prediction fidelity and policy generalization.
• UIUC & Amazon's CoDA: Data Engine and Policy Co-Evolution via Text2potential: Links a generative simulation engine and an action policy through a shared memory of environment code, failure diagnoses, and learning progress. VLM critic diagnoses are compiled into targeted environment generation programs and converted via Text2potential into staged potential functions for dense action-chunk value learning while preserving optimal policy invariance.

Authoritative Benchmarks & Empirical Comparisons

• VLA-Δ Sets 98.4% LIBERO Record at 2B Scale (arXiv:2610.21306): With a compact 2B backbone, VLA-Δ achieves a 98.4% average success rate on LIBERO (Spatial 98.8%, Object 99.8%, Goal 98.6%, Long 96.4%), outperforming ZR-0 (97.8%), openVLA-OFT 7B (97.1%), GR00T-N1.7 2B (97.0%), and π0.5 3B (96.9%). Under LIBERO-Plus language perturbations, VLA-Δ reaches 89.0% (+6.4 points over π0.5 at 82.6%) and beats π0.5 on RoboTwin 2.0 and real-robot benchmarks.
• OneWM Doubles π0.5 on RoboCasa365 (arXiv:2610.21095): On RoboCasa365, OneWM attains 36.68% overall success (Atomic-Seen 68.78%, Composite-Seen 30.25%), more than doubling π0.5 (16.90%) and beating GR00T N1.6 (21.90%) and GigaWorld-Policy (20.70%). Across seven OOD perturbation categories in LIBERO-Plus, OneWM achieves the highest overall score of 59.27% (vs. Fast-WAM at 49.70% and π0 at 53.60%).
• CoDA Lifts LIBERO-Pro from 1% to 76% (arXiv:2610.22362): Across four manipulation families, CoDA boosts OOD success from 5.3% after SFT to 27.6% (+9.5 points over the strongest closed-loop baseline). Over 15 co-evolution rounds on LIBERO-Pro, success jumps from 25% to 84% under position perturbations and from 1% to 76% under instruction perturbations.

Developer Practical Guide & Quickstart

Robotics developers can access code and checkpoints for VLA-Δ (https://github.com/ZJU-OmniAI/VLA-Delta), OneWM (https://github.com/Interactive-Active-AI/OneWM), and CoDA (https://plan-lab.github.io/coda) to train compact 2B–3B manipulation policies with physical state-transition supervision and closed-loop simulation curriculum generation.