SONIC-K1 is out: one policy that makes the Booster K1 follow thousands of human motions, including ones it never saw in training. It is NVIDIA’s SONIC, ported to the K1 and trained from scratch on about 10,000 mocap clips in about 100 GPU-hours, under 1% of SONIC’s compute.
In simulation it tracks 97.4% of 1,985 unseen motions and 93.6% of PHUMA, the external benchmark from the SONIC paper, with SONIC’s own success rule. The ONNX policy runs on a laptop in MuJoCo, no GPU needed.
Details and results: SONIC-K1 project and the release page. Next: the policy on a physical K1.
Training data includes Motion Data by Bones Studio; use of the underlying dataset is subject to the BONES Motion Capture Dataset License Agreement.