AdaMimic: Adaptive Motion Tracking for Humanoid Jumps and Racket Hits

One paper carries the name

AdaMimic is the method introduced in Towards Adaptable Humanoid Control via Adaptive Motion Tracking, arXiv 2510.14454, submitted on 16 October 2025. The authors are Tao Huang, Huayi Wang, Junli Ren, Kangning Yin, Zirui Wang, Xiao Chen, Feiyu Jia, Wentao Zhang, Junfeng Long, Jingbo Wang and Jiangmiao Pang, at Shanghai AI Laboratory and Shanghai Jiao Tong University, with Wang and Pang as equal advisors. The project page and repository describe it as an ICRA 2026 oral; arXiv lists no venue. [1] [2] [3]

A single reference motion, sparsified

The pipeline starts from one video of a human performing the skill. GVHMR recovers an SMPL motion, which is retargeted to the robot, sparsified into keyframes and lightly edited. Stage one trains a tracking policy with sparse global rewards and dense local rewards through separate critics. Stage two keeps that policy and trains two adapters, a phase adapter and a tracking adapter, which the project page describes as adjusting tracking speed and refining low-level actions. Training is PPO in Isaac Gym with 4,096 environments. [1] [2]

The point of the two-stage design is that the adapters make one tracked motion usable across a range of conditions, for example a longer jump or a different ball height, without retraining the base tracker. [1]

Hardware and rates on the G1

The robot is a 29-degree-of-freedom Unitree G1 with waist roll and pitch locked, so 27 joint dimensions are observed; the configs sit under g1_dof27. The policy runs at 50 Hz over a 500 Hz low-level loop, with odometry at 10 Hz from FAST-LIO on the onboard Livox MID360 lidar. [1] [3]

Six trials per setting is a small sample. The table shows that the skills ran on hardware; it does not estimate a success rate with any precision. [1]

Simulation numbers to read with care

Table III averages three evaluations of more than 10,000 episodes each. The AdaMimic default configuration reports 86.8 percent overall success, 30.3 mm local body-position error and 94.8 mm global error. The DeepMimic-Adapt baseline reports 85.1 percent, 33.3 mm and 133.5 mm. Other baselines are AMP-Style, AMP-Mimic and DeepMimic without adaptation. The global-error margin is the clearest difference; the success-rate margin is under two points. [1]

Code and licence

The InternRobotics repository was published with two commits on 16 and 17 October 2025 and no release. It builds on legged_gym, rsl_rl, HIMLoco, ASAP, AMP_for_hardware and GVHMR. The licence is CC BY-NC-SA 4.0, and the README states that commercial use is not allowed without explicit authorisation. The BibTeX in the README misspells one author's name, so copy the author list from arXiv. [3] [1]