Techniahqrobot | humanoid robots(@techniahqrobot):SkipVLA takes a pretty practical approach to VLA robot control. They tested it on a 6-DoF YAM arm. During free-space motion, the arm follows a classical planner. The VLA is called around grasping and placing. A small scorer looks at frozen vision-language features and picks the next 3D target pose. They did this without collecting extra demonstrations for that module. The real robot results are solid. With π0.5, three-block stacking went from 14/20 to 17/20 successful runs, while average completion time dropped from 47.62 s to 36.13 s. With MolmoAct2, placing two blocks in a box went from 18/20 to 20/20. Time dropped from 76.04 s to 41.66 s, and Jetson Thor compute energy fell from 3,237 J to 1,562 J. The largest measured time reduction was 59.6%. The stacking result caught my attention most. The robot finished more trials while calling the VLA less during the easy free-space parts of the motion. There is still a limitation. The switching rule currently depends on gripper open and close events, so the paper only tests pick-and-place style tasks.

Techniahqrobot | humanoid robots@techniahqrobot

2026.09.20 18:35

SkipVLA takes a pretty practical approach to VLA robot control. They tested it on a 6-DoF YAM arm. During free-space motion, the arm follows a classical planner. The VLA is called around grasping and placing. A small scorer looks at frozen vision-language features and picks the next 3D target pose. They did this without collecting extra demonstrations for that module. The real robot results are solid. With π0.5, three-block stacking went from 14/20 to 17/20 successful runs, while average completion time dropped from 47.62 s to 36.13 s. With MolmoAct2, placing two blocks in a box went from 18/20 to 20/20. Time dropped from 76.04 s to 41.66 s, and Jetson Thor compute energy fell from 3,237 J to 1,562 J. The largest measured time reduction was 59.6%. The stacking result caught my attention most. The robot finished more trials while calling the VLA less during the easy free-space parts of the motion. There is still a limitation. The switching rule currently depends on gripper open and close events, so the paper only tests pick-and-place style tasks.

Techniahqrobot | humanoid robots@techniahqrobot

2026.09.19 17:38

A Unitree G1 just showed how fast humanoid policies can fail when the environment moves outside their training range. New work from the University of Florida trained a G1 to perform roofing-style motions on pitched surfaces using human VR demonstrations. The human wore a PICO headset, two controllers and two ankle trackers. Those motions were retargeted to the 29-DoF G1. The team also measured the roof geometry and corrected foot support, hand clearance and body penetration before PPO training. In simulation, the robot performed nailgun positioning, hammering and lateral pushing. Work-clearance error reached 0.531 cm for nailgun positioning, 0.256 cm for hammering and 0.424 cm for pushing. All three tasks finished 3/3 successful runs. The 25° roof test is the result I keep coming back to. A policy trained on 9° and 17° slopes failed all 10 trials on an unseen 25° roof. Once 25° examples were added to training, the same setup completed 10/10 trials at 9°, 17° and 25°. The policies were also transferred to a physical G1. Base-frame motion error was 27.6 mm during uphill walking, 60.2 mm for nailgun positioning, 79.9 mm for hammering and 61.4 mm for bending. The hardware tests used a safety hoist. No nails were fired and hammer impact was not measured. Roof edges, shingles and weather were outside the test setup. That 10/10 to 0/10 drop at an unseen slope is a very concrete look at how narrow locomotion generalization can still be for humanoids.

Forward to community