Guanya Shi, an assistant professor at Carnegie Mellon’s Robotics Institute, presents his research on building agile, generalist robots at the RI seminar.
Shi defines generality as the range of tasks and environments a robot can handle, and agility as how quickly, precisely, and reliably it performs them. He contrasts specialized feats such as humanoid parkour with broad task systems such as RT-1, arguing that robotics still struggles to achieve both breadth and depth. He proposes combining physics-based structure with learning, and scaling online adaptation and optimization alongside offline training.
Shi describes H2O, which uses one RGB camera to teleoperate a Unitree H1 humanoid. Its pipeline filters human motions for physical feasibility in simulation, trains a deployable policy with reinforcement learning, then tracks eight body key points from camera-based pose estimation. OmniTool reduces the motion input to three points and uses supervised teacher-student training; Vocalo learns whole-body skills without human motion by breaking tasks into contact and task goals.
Shi says simulation-trained policies can be sensitive to simulator settings and can lose performance when transferred to the real world. He presents AnyCar, a transformer dynamics model pretrained on 100 million simulated samples and fine-tuned on about 20,000 real-world samples, then used with sampling-based MPC. He also describes hierarchical control for quadruped jumping and parkour, with reinforcement learning choosing high-level actions and a low-level controller tracking them, plus safety filters that can switch to recovery behavior.
Shi explains that sampling-based MPC can struggle with the discontinuous costs found in contact-rich tasks. His DPC method draws on diffusion models: it starts with broad, smooth action sampling and gradually reduces the sampling noise to refine the solution. For quadrupeds, it runs receding-horizon control at 50 Hz using full-order dynamics, without reduced-order models or preset contact sequences; he notes that the reported procedure ran offboard.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
Modi, a PhD researcher at Shanghai Innovation Institute, joins RoboPapers host Mike to explain EgoHumanoid, a method for combining human demonstrations with robot data to train humanoid loco-manipulation policies.
Ria, a USC assistant professor, and her student coauthors discuss their Humanoid Everyday dataset and humanoid manipulation research with the RoboPapers hosts.
NVIDIA Omniverse demonstrates how to use Mobility Gen in Isaac Sim to record robot trajectory data and render it from the robot’s camera perspective.
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.
