Unitree G1 humanoid uses IMU and joint states, trained on just 1 h of unlabeled data, achieving trajectories comparable to MAC-VO and SuperOdom. It matches visual and LiDAR odometry without a camera or LiDAR.
Key points
Training data: 1 h of unlabeled data.
Performance: trajectories comparable to MAC-VO (visual) and SuperOdom (LiDAR) with no camera or LiDAR.
Sources
One Scan is Enough: 85% success for language navigation
One Scan is Enough demonstrates a demonstration-free adaptation method that reaches an 85% real-robot success rate in language-guided navigation. The approach uses a single iPhone scan, Scan2Instruction to generate training episodes, and PRPT-KS for closed-form policy adaptation.
Key points
85% real-robot success across five real environments, versus 49% for a baseline trained on human demonstrations.
82% success in unscanned regions of the same buildings, versus 41% for the same baseline.
Sources
DYNAMIXEL XL430-W250-T – new compact smart actuator with multiple control modes
DYNAMIXEL XL430-W250-T is a compact smart actuator that supports position, extended position, velocity and PWM control. It weighs 65 g, measures 28.5 × 46.5 × 34 mm and includes a 4096-step encoder. The post invites building a TTL bus with this model.
Key points
Supports position, extended position, velocity and PWM control modes
4096-step encoder provides high-resolution feedback
Dimensions 28.5 × 46.5 × 34 mm; weight 65 g
Extended Position mode allows ±256 revolutions multi-turn control
Sources
- Original post
- DYNAMIXEL XL430-W250-T - Linked page
Joy dual-channel conversational dataset announced with 100,000 hours
Joy is a dual-channel conversational dataset introduced as Drop 1/15. It contains real, unscripted conversations between two people on separate tracks. The dataset totals 100,000 hours of audio. It aims to teach voice models conversational timing.
Key points
Dual-channel dataset of real, unscripted two-person conversations.
Contains 100,000 hours of audio.
Sources
DreamFAR released with real-time teleoperation
DreamFAR is a world model trained from scratch using only robot data, without internet pretraining. It supports real-time teleoperation and minutes of long-horizon rollout. It claims generalization across 1,000+ tasks.
Key points
Model trained from scratch using only robot data, without internet pretraining
Supports real-time teleoperation, minutes of long-horizon rollout, and generalization across 1,000+ tasks
Sources
ADT acquisition announced – essential service focus
ADT acquisition announced, describing home security as an essential service built on specialist expertise and earned trust. The acquiring company sees an opportunity to build on that foundation for decades, investing in it.
Key points
Acquisition of ADT announced.
Home security described as essential service with specialist expertise and earned trust.