Tech • AI • Robotics • Game
New Unitree G1 AI Robot Just Surpassed Human Skills
8/10TL;DR
Recent humanoid robotics results show fast imitation learning, improved whole-body control and more expressive interfaces, but many headline demos still lack independent validation and robust real-world metrics.
KEY POINTS
Berkeley’s Workhorse learns from sparse human data
UC Berkeley’s Hybrid Robotics Lab reported a system called Workhorse that teaches a Unitree G1 whole-body tasks from limited human demonstrations rather than robot teleoperation. A single demonstrator recorded 2.2 hours of box sorting, 1.2 hours of suitcase manipulation and 13 minutes of box catching using five HTC Vive trackers and a chest-mounted DJI Osmo Action 4 camera. The robot later used the same camera position and learned tasks with no direct joint-angle retargeting.
Robot actions include pushing, climbing and catching
In one showcase, the G1 pushed a 14-kilogram suitcase, knocked it flat with its leg, stepped up 0.39 meters onto it, reached a high shelf and jumped down. In another, it squatted, carried boxes to an upper shelf and kicked a third package into a lower shelf in one continuous run. The same platform also learned to catch thrown boxes after only 13 minutes of human examples.
Two-part architecture separates planning from balance
The system reduces motion to just five tracked points: chest, both wrists and both feet. A vision-based planner processes the last four camera frames plus recent motion history and predicts the next 1.16 seconds of those points at 5 hertz. A separate whole-body tracker, trained with reinforcement learning in simulation and running at 50 hertz, converts the short-horizon plan into joint commands while maintaining balance.
Robustness came from synthetic mistakes, not shove training
Researchers did not train the robot on examples of recovering from pushes, yet the machine still regained balance after kicks to the torso, pressure during a squat and even a box being pulled from its hands. That robustness came partly from deliberate error injection during training, including command shifts of up to 10 centimeters and 10 degrees and camera-image distortions. Human figures were also removed from training images and replaced with a rendered G1 to match what the robot would later see.
Simulation results are promising but real-world rates are missing
Across 100 simulated runs in a scanned room, the full system completed the three-box task in 77% of trials, falling to 64% when subjected to repeated 40-newton shoves. Removing planner error training dropped success to 20% and then 0% under disturbance. Despite strong demos, the paper did not report full real-world success rates, and all data came from one person in one room with task-specific models.
The method was transferred to another humanoid in simulation
The same approach was adapted to the larger Unitree H2 by remeasuring five mounting points and changing only the robot model in simulation. Reported simulated success was 83%, versus 77% for the G1, within the study’s stated error margin. That suggests some portability, though not yet a broad proof of generalization.
University of Washington focused on catching alone
A separate University of Washington project, Reflex, targeted dynamic catching on a G1 with a single onboard RGB-D camera. Training combined reinforcement learning for whole-body catching with learned prediction of box trajectories from delayed, partial observations. In simulation, camera-only control achieved 85.7% success versus 87.9% when given the box’s true state, but removing frame history cut performance by more than 35 points.
Real-world catching remains difficult
On physical hardware, Reflex stepped to intercept off-center throws and caught 8 of 20 side-directed attempts using packages weighing 0.5, 1 and 2 kilograms. The team has released code, paper and simulation replays, offering more outside scrutiny than many robotics demos. Even so, comprehensive real-world rates for easier chest-level throws were not provided.
Westlake Robotics is targeting continuous household work
Westlake Robotics presented WR1, a humanoid system designed to combine locomotion, posture and dexterous hand actions in one control stream. In demonstrations, the robot removed corn from a refrigerator, placed it in an air fryer, dragged a basket while picking toys off the floor and then straightened a bed in one uninterrupted sequence. The company also highlighted a public clothing-folding demo, including two robots coordinating on a pair of trousers.
Expressive robot faces are advancing faster than evidence of emotion
Viral attention also centered on an XROTDoll silicone robotic face that shifted through happy, angry, sad and surprised expressions in 15 seconds. The mechanics behind such systems are relatively clear, using small motors to drive eyebrows, eyelids and mouth corners, but what remains unclear is the control stack and whether interaction is autonomous or scripted. Across the sector, expressive faces are becoming more convincing, yet visible emotion display still says little about machine feeling or internal experience.
CONCLUSION
Humanoid robotics is moving beyond isolated arm motions toward integrated whole-body behavior, faster learning and more natural interfaces. The main gap is no longer impressive clips alone, but transparent real-world benchmarks that show how often these systems work outside carefully staged settings.
Ask a question