Experience-Based Feasibility-Aware Generative Adversarial Imitation from Observation under Embodiment Mismatch

Can an imitation-from-observation agent learn from state-only demonstrations that may contain robot-infeasible transitions by estimating feasibility from the robot's own online experience rather than from explicit dynamics models or pre-collected exploration data?

Robot-free demonstration interfaces record state trajectories without robot action labels, but human demonstrations can contain motions that are dynamically impossible or unsafe for a robot due to embodiment mismatch. Directly using these infeasible transitions in imitation can encourage the policy to reproduce motions that cause falls or failed grasping, and existing methods often rely on difficult-to-build whole-body dynamics models or separate prior exploration datasets.

Prior imitation-from-observation methods such as GAIfO and WGAIfO match state transition distributions but do not explicitly account for physical feasibility differences between demonstrator and learner. Model-based feasibility approaches use known or learned robot dynamics models to evaluate demonstrated motions, which is challenging for whole-body legged robots because of coupled dynamics, contacts, and balance constraints; weighting approaches for imperfect demonstrations typically need confidence labels or prior data.

EF-GAIfO learns a state-transition reconstruction model, a variational autoencoder, from the robot's own experience collected during policy learning. The reconstruction error of each demonstrated transition is mapped to a feasibility weight between 0 and 1 using adaptive thresholds. This weight is incorporated into the GAIfO discriminator loss, so high-feasibility demonstrations contribute more to policy updates. The robot's feasible region expands as the policy improves, allowing more demonstrations to become influential over training. For the quadruped task, a pretrained low-level locomotion policy is paired with a learned high-level policy that commands forward/lateral velocity, yaw, pitch, and gripper actions.

In point-mass simulation with 20 seeds and 200 evaluation trials per setting, EF-GAIfO achieved 90.0±30.0% success in a multi-speed setting versus 20.0±44.7% for WGAIfO and 15.0±35.7% for GAIfO, and 100.0±0.0% in a multi-path setting versus 20.0±41.0% and 50.0±50.0%. In quadruped simulation with 5 seeds and 100 trials per setting, EF-GAIfO reached 83.6–88.0% success across three fixed targets and 85.8±7.4% on randomized targets, while WGAIfO reached 26.4–39.0% and GAIfO 35.4–42.0%. On the real Unitree Go2, EF-GAIfO achieved 70–80% success across ten trials per setting, compared with 10–30% for WGAIfO and 10–20% for GAIfO.

The authors note that feasibility is currently estimated only from robot-experienced transitions, so environmental constraints such as terrain, obstacles, object placement, and contact conditions are not captured. Evaluation is limited to one quadruped robot, one object-reaching-and-grasping task, and small real-world trial counts of ten per setting. The paper does not report code release or compute requirements, and the simulation experiments use relatively simple point-mass and single-task settings; real-world transfer was tested only on the Go2 with the same task.