Where things stand
Grasping is getting good in controlled tests; fine finger work is still early. A UC San Diego system picked 10 household objects off a multi-level shelf with a Unitree G1 in 87.5% of 40 real trials, trained with no real demonstrations at all. Whole-body work is advancing too: a G1 pushed a cart carrying 30 kg. Long-duration reliability, tool use and in-hand manipulation on humanoids remain largely unproven.
The big shift is where the training data comes from. Teams are replacing hours of human teleoperation with synthetic data generated in simulation, and in at least one November test the simulation-trained robot beat robots trained on real demonstrations.
Humanoids are starting to use the whole body to grasp: moving the head to look, stepping closer and bending, not just reaching with an arm from a fixed spot.
Test objects are usually standard lab sets on clean shelves and tables. Results on cluttered, real-world shelves, soft or deformable items, and thousands of repetitions are still missing.
The papers
ForPC: Force-Aware Whole-Body Control for Precise Humanoid Cart Pushing · ZhenYu Zou et al. · arXiv 2611.08664 · Nov 5, 2026 · Preprint
- What they did
- Pushes a loaded cart along a route with both hands, estimating the forces from the robot's own joint readings instead of using force sensors.
- Result
- On a real G1, it pushed a cart carrying up to 30 kg at 0.5 m/s without extra tuning. In simulation it cut speed-tracking error by up to 68.2% and path error by up to 74.2% against a walking-only baseline.
- Caveat
- The baseline comparison is simulation-only; the real-robot test reports no success rate.
Checked against: Abstract. Results as reported by the authors; not peer-reviewed unless noted.
HuG: Learning Humanoid Whole-Body Affordance-Aware Grasping with Active Perception · Xuanbin Peng et al. (UC San Diego) · arXiv 2611.01659 · Nov 1, 2026 · Preprint
- What they did
- Teaches a humanoid to find an object on a shelf, turn its head to see it better, step and bend into position, and grasp it with dexterous hands. Trained only in simulation, with no real demonstrations.
- Result
- 87.5% success over 40 real trials with 10 household objects on a multi-level shelf, and 85% on objects it had not seen in training. The best baseline reached 52.5%; baselines trained on 70 human demonstrations per object and shelf level reached 32.5% to 50%.
- Caveat
- Lab test with standard benchmark objects. The model ran on an external desktop GPU connected by cable, not on the robot.
Checked against: Abstract and paper text. Results as reported by the authors; not peer-reviewed unless noted.
Moderate evidencePhysical robot and simulationNot named in the abstract
Anchor 1: Precise and Stable Humanoid Control for Whole-Body Manipulation · Archon Robotics, Caojun Wang et al. · arXiv 2611.00193 · Oct 30, 2026 · Preprint
- What they did
- A whole-body controller that copies a human operator's motion precisely and stably, to make teleoperated data collection faster.
- Result
- 30% to 80% lower tracking error than existing controllers in the authors' strictest real-world scoring, using fewer look-ahead inputs.
- Caveat
- The comparison controllers and the robot are not named in the abstract. This is teleoperation, not autonomy.
Checked against: Abstract. Results as reported by the authors; not peer-reviewed unless noted.