NVIDIA's continual learning system autonomously writes and refines robot control programs via code-as-policy. Surpasses prior methods by up to 77% on LIBERO-Pro and enables zero-shot generalization with 31% success vs 4% for prior methods.
New 33B total parameter MoE model specialized for agentic coding with native interleaved thinking/tool-call reasoning, 262K context, and SWA + global attention. Scores 70.9% SWE-bench Verified and runs locally on Mac with 36GB unified memory.
NVIDIA's official NVFP4 quantized Qwen3.6-27B shrinks to ~1/3 BF16 size while matching FP8 quality (86.3 MMLU-Pro) and preserving 262K context. LMSYS added Day-0 SGLang support — makes a strong 27B model far more runnable on consumer hardware.
P2R decouples perception and reasoning into a two-stage process trained with Perception-Reasoning Alternating GRPO. On Qwen3-VL-4B achieves 93.2% on V-Star, 81.9% on HR-Bench-4K.
Fits a causal model on 512 runs of Qwen2.5-0.5B to estimate Conditional Average Treatment Effect, then extrapolates optimal mixtures for 800K data pools and 7B models. Outperforms RegMix across downstream tasks.
AMVL resolves train-inference mismatch in multimodal continuous reasoning using bidirectional calibration with forward and reverse KL divergences. Gains up to +32.00 on individual reasoning tasks.
ELDR is an expert-locality-aware decode router for prefill-decode disaggregated MoE serving that predicts expert activations and routes accordingly. On up to 40 GPUs reduces median TPOT by 5.9-13.9% over load-balancing baselines.
Domain Arithmetic (DART) adapts VLAs to environmental shifts like camera pose and embodiment changes using weight vector arithmetic with subspace alignment. Accepted at ECCV 2026.
Alibaba affiliate Ant Group led the round, marking its 12th humanoid robotics investment. Zeroth claims orders exceeding 30,000 units and 600% revenue growth YoY in H1 2026, with plans for home robots and Alipay payment integration.
The Humanoid Application Center opened in Schiedam, bringing together firms, researchers and technicians to close Europe's robotics gap with the US and China, helping companies deploy humanoid robots in industrial settings.
Nvidia will let fast-growing startups swap access to compute power for a slice of future profits. Sharon AI will deploy up to 40,000 Nvidia GPUs, and Firmus Technologies is building a 360MW data center in Indonesia housing up to 170,000 GPUs.
Beijing-based Striding AI integrates foundation models, perception, control systems, and real-world action data. Human-in-the-loop RL improved task success rates by up to 3x in internal testing, targeting retail, logistics, healthcare, and agriculture.
Z.AI launched ZCode, a development platform and harness for the GLM-5.2 model enabling coding, agentic workflows, and model interaction. Hit #5 on Hacker News with 417 points.
Meta's Pocket lets users generate interactive mini-games and apps using plain-text AI prompts. Built on acquired Gizmo team's tech, it offers a social feed of AI-generated 'gizmos' with motion, camera, and audio controls.
Open-source tool creates a deterministic architecture graph from codebases, designed to help both human developers and AI agents understand project structure and dependencies.
A secondary tender offer values the AI voice startup at ~$22B, up from $11B in February's $500M Series D, reflecting surging demand for AI voice generation technology.
Growth-stage funding for AI-powered homebuilding design and construction platform, showing AI's expansion into the construction sector.
Early-stage AI platform automating data center site selection and due diligence, tapping into the booming demand for AI infrastructure real estate.
Source:
