MimicAgent

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

We present MimicAgent, a prompt-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. Our key observation is that it is far easier for a human – and by association, an LLM – to generate reference motions than to shape reward functions. Our hypothesis is motivated by the success of example-guided RL for humanoids, which exploits large-scale motion capture datasets as references for training locomotion policies. Unlike humanoids, quadrupeds lack such reference motion data. Towards this end, we propose MimicAgent, an agentic harness that, given a skill prompt, generates quadruped reference trajectories with coding agents. These coarse reference trajectories are then used to train example-guided RL policies that are deployable in simulation and in the real-world. Notably, we find that when prompting Claude Fable 5.1 within our agentic harness, 87% of prompts yield semantically aligned reference trajectories.

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.

Preliminary results show that MimicAgent can be adapted for humanoid locomotion on the Unitree G1. Given the humanoid's model and configuration as context, the same prompt-to-trajectory framework generates diverse whole-body skills, from acrobatics to climbing and interaction with objects. Each clip shows the G1 reference motion (in green).

MimicAgent policies transfer from simulation to the physical robot. The clips below show several skills executed on real hardware, including dynamic side flips, back flips, and a four-leg to handstand spin transition.

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: Quadruped Skills via Prompt-to-Trajectory Generation},
  author  = {Lucky Kant Nayak and Narayanan Palghat Parameswaran and Neehar Peri and Deva Ramanan},
  journal = {arXiv preprint arXiv:2609.24145},
  year    = {2026}
}