MimicAgent

We visualize never-before-seen skating skills on two legs generated by MimicAgent.

We present MimicAgent, a text-to-trajectory generation framework for learning dynamic quadruped skills. Although reward shaping is extensively used when training quadruped policies, navigating the resulting reward landscape is notoriously difficult, requiring hours of ``graduate student descent''. Eureka attempts to automate reward design with LLMs, but we find that it struggles to generalize across diverse skills and morphologies. Motivated by the success of example-guided RL for humanoids, we revisit skill learning from demonstrations for quadrupeds. Unlike humanoids, which can exploit large-scale motion capture datasets for learning, quadrupeds lack such reference motion data. We make the observation that manually keyframing quadruped reference motions can be more intuitive than reward shaping; in particular, we find that rather coarse and even dynamically-infeasible motions can still be effective reference targets for example-guided RL. However, manual keyframing is still too cumbersome to create large-scale skill libraries. To address this challenge, we propose an LLM-based pipeline that generates kinematically feasible quadruped trajectories for diverse skills. Although these trajectories are not dynamically feasible, we show that they are sufficient to train successful policies. Across all evaluated skills, human raters often prefer policies generated by MimicAgent over those produced by Eureka.

We compare MimicAgent against two state-of-the-art baselines (Eureka and Manual Keyframing) across seven skills: Trot, Bound, Side Flip, Front Flip, Aerial Crossover (AC), Crab Diagonal Scuttle (CDS), and Reverberating Yaw Pulse (RYP). Trot and Bound run on the Unitree Go2; the remaining five skills run on its wheeled-legged variant, Go2-W. Each video demonstrates the performance of different methods on the same skill.

User study (n = 57). Mean human preference rating (1–5, higher is better) for policies from human-annotated keyframes, Eureka, and MimicAgent. Bold marks the top-rated method for each skill. Users prefer MimicAgent for Trot, Side Flip, Front Flip, Aerial Crossover, and Crab Diagonal Scuttle; human keyframes for Bound (where all methods score low); and Eureka for Reverberating Yaw Pulse.

MimicAgent learns and executes a wide range of locomotion skills directly from text descriptions. The green robot shows the reference trajectory generated by MimicAgent, and the blue robot shows the learned policy.

MimicAgent is not limited to quadrupeds, the same text-to-trajectory pipeline generates full-body humanoid motions. Shown here on the SMPL body model are everyday motions that already appear in standard motion-capture datasets; MimicAgent reproduces them directly from a text prompt.

In contrast to the everyday motions above, these are novel, difficult skills — dynamic acrobatics and contact-rich interaction with objects and terrain — many of which rarely or never appear in motion-capture datasets. MimicAgent synthesizes each one from a text prompt alone, going beyond what standard mocap corpora contain. Each clip shows the humanoid reference motion (in green) produced by MimicAgent.

MimicAgent policies transfer from simulation to the physical robot. The clips below show several skills executed on real hardware.

We distill the trained teacher policy into a deployable student policy that runs from onboard observations alone. Below, all three tracks execute the same four-leg-to-handstand spin transition side by side, the student stays in near-lockstep with both the teacher and the reference, showing that distillation preserves this difficult, dynamic skill.

If you find our work useful, please cite:

@article{nayak2026mimicagent,
  title   = {MimicAgent: Learning Quadruped Skills via Text-to Trajectory Generation},
  author  = {Lucky Kant Nayak and Narayanan Palghat Parameswaran and Neehar Peri and Deva Ramanan},
  journal = {Conference/Journal Name},
  year    = {2026}
}