Ria, a USC assistant professor, and her student coauthors discuss their Humanoid Everyday dataset and humanoid manipulation research with the RoboPapers hosts.
The team collected more than 10,300 trajectories across 260 tasks, totaling over three million frames from Unitree G1 and H1 robots. They synchronized RGB, depth, LiDAR, tactile, and robot-state streams at 30 Hz, and built a viewer for exploring 50 example tasks. The dataset covers seven areas, including tool use, locomotion, and high-precision manipulation.
The team evaluated imitation learning and vision-language-action policies on tasks such as folding a towel, inserting a flower into a vase, and handling a door. Pretrained VLA models generally did better than traditional imitation methods; GR00T N1.5 averaged 51% success, while the most difficult vase insertion reached only about 10% with OpenVLA. The team said roughly 20 to 30% of early recordings were discarded because of timing, frame, or camera-view problems.
The project took nearly nine months, involved several students and additional data collectors, and resulted in two damaged G1 robots that Unitree helped repair. The published data contains about 30 hours of recordings, while the team estimated roughly 151 hours of teleoperation for the tasks alone, before counting planning, quality checks, and hardware repairs. Operators needed practice adapting to the robot’s fixed neck and narrow camera view.
The researchers found that pretraining on Humanoid Everyday improved later task tuning: for one insertion task, π0.5 had no successes without that prior and four successes in ten trials with it. They also described a cloud evaluation platform that sends robot images and state to a researcher’s client and returns inferred actions. The team hopes to use it for controlled, comparable evaluations and possibly a challenge.
The current dataset mostly uses simple lower-body movements, and the team said VLA models struggle with more complex motions such as crouching or sitting. They want to explore whole-body policies, action tokenization, and human egocentric video as more scalable training data. The researchers also argued that teleoperation remains useful for aligning policies to robot actions, even if other data sources are needed to scale collection.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Lee Yi, an assistant professor at Tsinghua University and coauthor of R2S2, discusses whole-body humanoid reaching and teleoperation with the RoboPapers hosts.
PRO ROBOTS’ main presenter covers humanoid, service, industrial, and healthcare robots at China’s 2025 World Robot Conference, with demonstrations and interviews from exhibitors.
Sergey, host at Supercar Blondie, explores Unitree Robotics’ quadruped and humanoid robots at the company’s factory in China with Unitree representative Yolanda.
Jen Jang of Galbot presents OpenWBT, an open-source VR teleoperation system for whole-body humanoid control in simulated and real environments.
Aadeel Akhtar of PSYONIC presents at Humanoids Summit on bionic hands that serve prosthetic users and help robots learn manipulation tasks.
Light Origins’ presenter explains how Light-O1 uses human video to train robot action models and demonstrates it on a Lightwheel setup and a Unitree G1.
