Sim/digital twins & embodied

Gamma-World (24 FPS); Qwen-VLA; NVIDIA Cosmos 3; World-Language-Action (40ms); Flash-WAM (23x); OASIS (zero-shot sim-to-real). ASPIRE autonomous skill discovery (31% zero-shot on LIBERO-Pro Long). Embodied.cpp portable runtime. VLA-Corrector. Forward Networks digital twin for network risk/autonomy. GigaWorld-1: systematic study of world models for robot policy evaluation with WMBench and 324K simulated rollouts; key insight: long-horizon action-faithful rollout consistency matters more than short-term visual realism. WorldSample (world model for robotics). EVA-Client: unified data collection/inference/deployment framework for embodied policies on real robots (decoupled architecture, inspectable workflows). InternVLA-A1.5: unifying understanding, latent foresight, and action for compositional generalization (avoids pixel-level future prediction, strong sim/real results). Digital Twin Synchronization Over Mobile Embodied AI (MEAN framework with five-stage closed-loop workflow, hierarchical optimization, semantic compression under bandwidth constraints). Generating Synthetic User Populations via World Models — new paper on realistic synthetic populations for simulation. Synthetic Data vs Real Data for Enterprise AI Training (2026) — practical hybrid strategy guide. RynnWorld-Teleop — action-conditioned world model for digital teleoperation; 40+ FPS real-time generation, zero-shot Sim2Real; decouples data collection from physical hardware. RynnWorld-4D — 4D embodied world model using RGB-DF representation; tri-branch architecture with cross-modal attention and 3D RoPE; 254M-frame dataset; policy head bypasses multi-step denoising for closed-loop action. SIEVE — structure-aware data selection for VLA imitation learning; outperforms full-data training with 50% data and 50% steps. LeCun world models discussion — 1.5-hour deep technical discussion on world models; reinforces simulation priority. MIRA — 5B multiplayer world model trained on 10k hours of Rocket League data, playable at 20fps on single GPU; interactive neural world model without physics engine; collaboration between General Intuition, Kyutai, Epic Games. MoE Video Pretraining for Embodied — scaling MoE video pretraining with robot-oriented data and multi-dimensional reward system; open-source, addresses domain mismatch in video models for robotics. RoboDojo — unified sim-and-real benchmark for generalist robot manipulation policies; 42 sim + 18 real tasks, evaluating 30 policies across generalization, memory, precision, long-horizon, open-vocabulary; reproducible real-world evaluation with remote cloud access. Physical AI Compute cost analysis — robots need to generate own training data via simulation, shifting cost from single training run to ongoing simulation, multimodal training, and continuous inference; concrete numbers (780K trajectories in 11 hours) useful for infrastructure planning. New: Multi-agent robotic control with onboard VLMs (3B params) using Megamind orchestration, fine-tuning improved inspection accuracy 76.7%→91.5%, open-source sim environment. New: Automating Design of Embodied Agent Architectures — AgentCanvas runtime, KDLoop search, 3x4 evaluation matrix across VLN/EQA/manipulation, honest failure mode reporting. New: MALLM-GAN — multi-agent LLM emulating GAN for tabular data synthesis in low-data regimes, code available. New: Mercor acquires Deeptune for simulated training environments for AI agents — $20B valuation, $2B ARR, validates RL-based agent training infrastructure as critical bottleneck. New: AgentWorldBench — benchmark for language world models across 7 domains with multi-dimensional scoring, based on real trajectories from established benchmarks. New: Open-ended multi-agent autocurricula via visual inspection of policy videos using VLMs — addresses limitation of scalar scores in curriculum generation; tested on SMAC. New: SocialDropout — dynamic agent dropout for social simulation using RL-based selection; reduces computational cost while preserving behavioral realism; practical optimization for large-scale LLM-based simulations. New: Sakana AI revisits Picbreeder with modern VLMs and LLM agents for open-ended exploration (GECCO2026) — explores computational creativity mechanics (serendipity, novelty search); relevant to open-ended autocurricula and world models. New: Practical piece on virtual gyms for robotics deployment — covers sim-to-real gap, selective fidelity, synthetic data workflow, lifecycle integration; Toyota forklift perception example (99.5% precision); 30-50% commissioning time reduction. New: Meta-Twins — system-of-systems digital twin for data infrastructure itself; encodes pipeline intent as relational metadata with daemon enforcement; SQL-native control plane for AI-assisted updates; novel application of digital twin concepts to autonomous data pipelines. New: Jeff Dean tweet on fine-grained subtask annotations improving VLA performance — SOTA vision+proprioception model achieves 93.1 F1@50 on REASSEMBLE and 98.6 on Amazon Robotics blade insertion. New: NVIDIA RoboLab evaluation guide — Clopper-Pearson analysis shows 15x more rollouts needed for statistical significance; challenges small-N binary success reporting; practical for deployment decisions. New: MCP server for Unity digital twins — open-source, makes any static C# method an agent-callable tool, lowers barrier for agent-simulation integration. New: Flow-ERD — agent-type aware flow matching with entropy-regularized distillation for diverse traffic simulation; Pareto dominance over prior methods; log-free diversity metric. New: Nvidia GR00T 1.7 VLA model and Cosmos 3 world model open-sourced on Hugging Face, with 32K hours real demos and 8K sim data, integrated with LeRobot. New: Xiaomi open-sources 38B robotics world model with 82.9x inference speedup via FlashAR+, 26.3pp improvement on π0.5 OOD success rate. New: SPEAR — photorealistic simulator connecting Python to Unreal Engine, 14K+ functions, 73fps rendering, ground truth modalities, deterministic execution. New: 4D embodied world model (tri-branch DiT, cross-modal attention, 250M+ RGB frames with depth/flow) — advances simulation fidelity for embodied agents. New: Agentic Economic Modeling — LLM-generated synthetic choices corrected with small human samples; practical for econometric inference and A/B testing. New: RoboTTT — scales visuomotor context to 8K timesteps via test-time training, enabling one-shot imitation and long-horizon tasks; 87% improvement over single-step baseline, 62% gain from scaling context; concrete real-robot results. New: NVIDIA shrinks Blackwell to Jetson Thor T3000/T2000 for mass-market robots, Cosmos 3 Edge (4B params) for on-device world model inference; T3000 lacks MIG partitioning — important for multi-model isolation; one-day fine-tuning claim for sim-to-real. New: Step-Level Preference Learning for Generative Agents (SimPref) — step-level DPO for social simulation; 57K preference dataset; improves simulation fidelity; open-source. New: Causal graph guided MARL for multi-agent cooperation — counterfactual interventions for dynamic communication graph; generalizable approach. New: Japan Cosmos Coalition — 22 industrial firms join NVIDIA's Cosmos Coalition for physical AI deployment; collaborative control platform with Fujitsu, FANUC, etc. New: NVIDIA Omniverse Agent Toolkit — libraries for automated simulation world creation (RTX sensor sim, GPU physics, SimReady validation); open-source, reduces manual simulation setup. New: AOUSD drives 3D data interoperability for agentic AI — new members (ByteDance, Huawei, Physicl, Unity), ISO certification, OpenUSD standard for 3D environments. New: Masked Visual Actions for Unified World Modeling — pixel-space action representation enabling forward/inverse dynamics from single model; 15 hours finetuning. New: Stress-Testing Grounding Models on Synthetic Data — KubriCount benchmark; LocateAnything-3B 0.00 recall on 18 categories; practical FiftyOne workflow.