Newsfeed
Real-time intelligence on humanoid robotics funding, launches, research, and policy.
AMD ROCm Powers VLA Humanoid Training Without CUDA
AMD ROCm pipeline trains and deploys VLA manipulation policies on humanoids without NVIDIA CUDA.
Read →ARCHITECT Beats VLAs With Modular Code Synthesis
ARCHITECT outperforms VLA models on long-horizon tasks by treating policy learning as interactive code synthesis.
Read →FutureRTC Fixes VLA Timing Gaps Without Retraining
FutureRTC solves VLA inter-chunk discontinuities via anticipatory observation prediction, no policy retraining required.
Read →Tactile Glove Artifacts Cut 18.3% by Pose-Aware Model
New pose-aware framework cuts tactile glove minimum detectable force by up to 18.3% across 3 glove designs and 15 users.
Read →WCM Hits 73.8% Success Rate on 9 HRI Tasks
WCM's SLAK architecture achieves 73.8% success across 9 real-world HRI tasks, with a human-in-the-loop teaching mode for long-horizon tasks.
Read →DynaMAC Beats Baselines by 35pts on Bimanual Tasks
DynaMAC outperforms leading baselines by 35+ points on bimanual tasks while needing 20x fewer training samples.
Read →Pigey Closes the Orchestration Gap 4x on LIBERO-PRO
New 'Pigey' orchestrator lifts frozen VLA policies from 12.8% to 53.3% on LIBERO-PRO with no fine-tuning.
Read →Synthetic Video Trains Humanoids with Zero Real Data
New framework uses generative AI to convert text prompts into motion demos, training humanoids across four tasks with zero real-world data.
Read →China Targets 10,000 Humanoid Units in Factories by Year-End
China deploys humanoid robots in garment and electronics factories, targeting 10,000 units by end of 2026.
Read →IEEE Humanoids 2026 Makes Labor Displacement Its Central Theme
IEEE Humanoids 2026 names labor displacement its defining theme as 1,200 researchers convene in Santa Clara this December.
Read →Optimus Targets 2,500 Units/Week as Musk Flags Supply Chain Risk
Tesla targets 2,500 Optimus units/week by year-end, but Musk calls it the hardest product Tesla has ever tried to manufacture.
Read →AXIS Dataset: 207 Tasks and 50K Trajectories for VLA Training
AXIS delivers 207 manipulation tasks and 50K+ trajectories via browser teleoperation, boosting π0.5 success rate by 5.8%.
Read →MoE VLAs Learn Task Primitives Without Explicit Hierarchy
New arXiv paper shows MoE action heads in VLAs spontaneously learn reusable skill primitives without any predefined task hierarchy.
Read →Holiday Robotics Raises $105M for FRIDAY Wheeled Humanoid
Seoul-based Holiday Robotics closes $105M Series A for FRIDAY, a wheeled humanoid with 64 DOF and 40 dedicated to its hands.
Read →Clinical Pathways as Runtime Safety Specs for Hospital AI
Researchers propose using clinical pathways as enforceable runtime safety constraints for assistive robots in hospital wards.
Read →DEED VLA Framework Cuts Humanoid Retail Training to One GPU
DEED framework turns failed naive fine-tuning into competent chip-restocking on a Unitree G1-Edu using one GPU and GR00T N1.6.
Read →Extreme-RGMT Solves Humanoid Continual Learning Tradeoff
Extreme-RGMT framework tackles the generalist-specialist tradeoff in humanoid motion control using two-stage continual learning.
Read →ModPack Brings Hardware-Agnostic Teleoperation to Humanoids
Columbia researchers open-source ModPack, a wearable teleoperation system that works across robot platforms for bimanual mobile manipulation.
Read →YAHMP Framework Isolates Key Factors in Humanoid Motion Tracking
HUCEBOT researchers release YAHMP, an open-source framework that isolates which design choices actually move the needle in humanoid motion tracking.
Read →FARO Combines LLM Planning with Kino-Dynamic Optimization
FARO uses LLM-guided contact planning and kino-dynamic feasibility checking to generate real-world humanoid loco-manipulation trajectories.
Read →