Video-to-robot pipelines for quadrupedal imitation insert a learned lifting stage, either a monocular depth model or a per-clip 3D reconstruction, to recover the one axis a side-view camera cannot see. We present CopyQuad, which omits that stage and pins the missing axis to the robot's own stance width instead. The setting is side-view footage of sagittally dominated motion, where limb motion is almost entirely in the image plane. A frozen 2D keypoint detector and a fixed analytic retargeting then turn a single uncalibrated clip into a robot reference whose foot targets are reachable by construction, so no per-clip trajectory optimization is needed. One phase-indexed reward structure, whose only per-behavior freedom is three scalars from a shared sweep, learns a dog walk, a horse walk, a horse trot and two non-periodic gestures, all deployed zero-shot on hardware. We then measure what omitting the axis costs. In a training-free rigidity test, a video depth model varies the distance between the animal's hips by 40 to 107% of its own mean and a single-image metric model by 22 to 58%, though that distance cannot move. Motion capture puts the true lateral paw travel of a dog walk at 4.5 cm peak to peak, a mean deviation near 1.4 cm against the 3.5 cm of lateral foot error our policies already leave, so the discarded axis sits below the pipeline's resolution. Depth repaired to feasibility never beats the constant, and the unrepaired estimates put the robot on the ground.
Figures & tables
Fig. 1 : CopyQuad. A single monocular clip becomes a robot-ready reference through one frozen 2D keypoint detector followed by a chain of fixed analytic steps, namely sagittal-plane lifting with the lateral axis pinned to the robot’s stance width, kinematic retargeting, and root recovery. A phase-indexed policy πψ trained in simulation tracks that reference and is deployed zero-shot on hardware. Only the reference itself and three reward scalars change between behaviors.
Fig. 2 : The reference quantities. Left : the stance tracks yknom=±w/2 ( dashed ) that Eq. 1 pins the lateral axis to, with the base x^ and speeds vx , vy . Middle : the robot height Hr , a foot pℓ and the stance sweep dˉ . Right : the detector keypoints , all that is read from pixels, and the spans La , Ha that Eq. 2 matches to Lr , Hr .
Fig. 3 : Source video → Isaac Lab → real robot for the two gesture clips. All four feet are tracked, but only one front and one hind are ringed, ∙ front and ∙ hind, since left and right overlap in a side view.
Fig. 4 : Source video → Isaac Lab → real robot for the two gesture clips. The gestures are read from the near-side pair alone, so only two feet are ringed, ∙ front and ∙ hind.
Source clip
Dog walk
Horse walk
Horse trot
Dog bow
Dog give-paw
fps, length (s)
30, 21.4
25, 21
30, 36
50, 7
30, 8
Simulation
vˉ^→vˉ (m/s)
0.29→0.29
0.48→0.55
0.58→0.58
0.00→0.00
0.00→0.00
eq per-clip (deg) ↓
14.3±1.4
13.3±1.7
11.6±1.3
4.7±0.3
8.3±4.4
eq shared (deg) ↓
18.1±1.6
15.7±2.1
13.4±1.9
4.2±0.4
4.6±0.7
ep (cm) ↓
13.7±1.8
15.0±2.6
15.5±2.0
7.4±0.4
11.8±6.8
ep fore-aft / lateral / vert. (cm)
9.8/3.5/3.7
10.5/3.8/4.0
11.0/3.9/4.2
—
—
TABLE I : All five behaviors on ANYmal-D, from the source clip through simulation to zero-shot hardware. eq per-clip and eq shared differ only in the reward weights, namely three per-clip scalars or one fixed vector for all five, and both are evaluated against the same full-method reference, so the per-clip row is the full bar of Fig. 5 . A dash marks a quantity undefined for a stationary gesture.
eqref (deg) ↓
vˉ (m/s)
Surv. (s)
CopyQuad (ours)
14.3±1.4
0.29
19.9±0.2
Replacing the lifting stage
SLoMo [ 8 ] †
33.6−36.6
≈0
—
SLoMo+physics refinement [ 8 , 26 ] †
19.5±1.1
0.22±0.1
20.0
Replacing the objective
AMP [ 27 ] †
—
0.03
—
TABLE II : Alternatives on the dog walk. Each row replaces one stage and leaves the rest untouched, the upper group the lifting stage and the lower group the objective. All use eqref against the full method’s reference, survival capped at 20 s. Dashes are undefined quantities. † marks the need to an extra conversion (SLoMo) or a data regime we do not supply (AMP).
VDA (relative)
Depth Pro (metric)
Clip
∣Δzc∣
CV
∣Δzc∣ (cm) / CV
sign
Dog walk
2.42
40%
9 / 22%
98%
Horse walk
0.00
—
6 / 35%
86%
Horse trot
0.01
—
4 / 58%
61%
Dog bow
1.68
75%
11 / 28%
95%
Give-paw
0.94
107%
10 / 31%
94%
TABLE III : The rigidity test. Implied left-right hip separation ∣Δzc∣ along the camera’s depth axis, subscripted to distinguish it from the robot’s vertical z , under a relative video model (VDA [ 28 ] , model units) and a single-image metric model (Depth Pro [ 29 ] , cm). CV is blank on the horse clips, where VDA leaves no separation to take a ratio of.
Fig. 5 : Retargeting ablation on the dog walk and the horse trot. Bars are eqref against the full method’s reference. Every variant is evaluated on one fixed target, with whiskers ±1 std over seeds. Color marks what the policy did, not how well it tracked. Blue survives the horizon and locomotes. Red falls before 15 s, so its eq is post-fall accumulation rather than tracking, with survival annotated. Gray stays up but never locomotes, so a low eq is a stationary robot.
MoCap route
Video route (ours)
eq (deg) ↓
9.5±1.0
12.5±1.4
ep (cm) ↓
10.5±1.5
13.5±1.8
Surv. (s, cap 20)
20.0
20.0
TABLE IV : What video costs against MoCap, on one dog walk taken from markers and from a side-view rendering of the same motion. Both routes share the retargeter, reward and budget, and both are evaluated against the marker-derived reference.
Quadruped robots have achieved remarkable locomotion, yet their behavioral repertoire remains confined to a few gaits--far from the expressive, companion-like presence long envisioned for them. Attempts to import the humanoid recipe of large-scale motion data have inherited one tacit assumption: that robot motion must first pass through an animal body, making data collection dependent on cooperative animals, reconstruction fragile across species, and retargeting ill-posed across incompatible morphologies. We propose Uni-Mo, a fully automated pipeline that removes the animal from the loop by reframing data scarcity as a generation problem: an LLM proposes motion prompts, a video diffusion model synthesizes the corresponding robot behaviors, and the generated videos are lifted into 3D reference trajectories used to train tracking policies deployed on a real Unitree Go2. To make naively-drifting generations reliably extractable, we introduce an Identity Consistency Loss that enforces appearance coherence across frames. We release Quad-Imaginarium at https://github.com/Amap-Robotics/Quad-Imaginarium, the resulting open-source dataset of 7,488 language-annotated quadruped motions (18.5 hours) spanning acrobatic and performative behaviors. We validate 392 randomly sampled motions on a real Unitree Go2 with a 96.7% deployment success rate, complemented by a 97.6% success rate across the full dataset in simulation.
Quadruped locomotion provides a natural setting for understanding when model-free learning can outperform model-based control design, by exploiting data patterns to bypass the difficulty of optimizing over discrete contacts and the combinatorial explosion of mode changes. We give a principled analysis of why imitation learning with quadrupeds can be inherently effective in a small data regime, based on the structure of its limit cycles, Poincaré return maps, and local numerical properties of neural networks. The understanding motivates a new imitation learning method that regulates the alignment between variations in a latent space and those over the output actions. Hardware experiments confirm that a few seconds of demonstration is sufficient to train various locomotion policies from scratch entirely offline with reasonable robustness.
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.
Lucky Kant Nayak, Narayanan Palghat Parameswaran, Neehar Peri +1