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

Yoshiki Takebayashi; Giovanni Perantoni; Hikaru Sasaki; Matteo Saveriano; Takamitsu Matsubara

cs.ROOct 1, 2026

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

Authors: Yoshiki Takebayashi, Giovanni Perantoni, Hikaru Sasaki, Matteo Saveriano, Takamitsu Matsubara

Organizations: Division of Information Science, Graduate School of Information Science, Nara Institute of Science and Technology (NAIST), Nara, Japan · Department of Industrial Engineering of the University of Trento, Trento, Italy

Abstract

With the increasing use of robot-free demonstration interfaces that provide state trajectories without action labels, imitation from observation has become a promising approach for learning robot behaviors from human demonstrations. However, due to differences in embodiment and dynamics between humans and robots, demonstrated human motions may not be feasible for the robot, potentially degrading policy performance. In this study, we propose Experience-Based Feasibility-Aware Generative Adversarial Imitation from Observation (EF-GAIfO), which estimates the feasibility of state-only demonstrations from the robot's own experience rather than relying on explicit dynamics models or large prior exploration datasets. A key feature of EF-GAIfO is that the notion of feasibility evolves with policy learning: as the policy improves and the robot experiences a broader range of state transitions, the feasible region is progressively expanded, allowing additional demonstrations to be incorporated into learning. This enables feasibility-aware imitation that adapts to the current stage of policy learning, rather than relying on a pre-designed feasibility criterion. We validate the effectiveness of EF-GAIfO on a locomotion task in simulation and on a real quadruped robot performing a object-reaching-and-grasping task.

Figures & tables

Fig. 1 : Evolution of feasibility estimation during policy learning. Early in learning, limited robot experience yields high feasibility estimates only for demonstration transitions near the experienced region. As experience accumulates, the covered region expands, allowing more demonstration transitions to receive high feasibility estimates. This progressively reduces the influence of infeasible demonstrations during imitation learning.

Fig. 2 : Overview of EF-GAIfO. Robot experience is used to estimate the feasibility of state-only demonstration transitions, and the estimated feasibility weights are incorporated into discriminator training for feasibility-aware policy learning.

Panel of Figure 3 — Fig. 3 : Demonstration trajectories in the point-mass environment for (a) the multi-speed setting and (b) the multi-path setting. Red and blue trajectories represent infeasible and feasible demonstrations, respectively. Color intensity indicates execution speed, with lighter colors denoting slower motions.

Setting	Method	Success rate ( $\uparrow$ )	Hausdorff distance ( $\downarrow$ )
	EF-GAIfO	$\mathbf{90.0\pm 30.0\%}$	$\mathbf{0.884\pm 1.72}$
Multi-speed	WGAIfO	$20.0\pm 44.7\%$	$2.629\pm 2.694$
	GAIfO	$15.0\pm 35.7\%$	$4.877\pm 2.396$
	EF-GAIfO	$\mathbf{100.0\pm 0.0\%}$	$\mathbf{0.163\pm 0.075}$
Multi-path	WGAIfO	$20.0\pm 41.0\%$	$1.742\pm 2.343$
	GAIfO	$50.0\pm 50.0\%$	$2.828\pm 2.57$

TABLE I : Mean task success rates and Hausdorff distances in the point-mass experiment. The Hausdorff distance measures the discrepancy between the trajectories generated by the learned policy and the feasible demonstration trajectories. Values are reported as the mean

\pm

standard deviation of the learned policy over 20 random seeds. Each learned policy is evaluated over ten episodes.

Setting	Demonstration	Iter.	EF-GAIfO	WGAIfO
	Feasible ( $\uparrow$ )	100	0.142 $\pm$ 0.032	0.450 $\pm$ 0.039
Multi-	Feasible ( $\uparrow$ )	3000	0.924 $\pm$ 0.105	0.402 $\pm$ 0.473
speed	Infeasible ( $\downarrow$ )	100	0.016 $\pm$ 0.030	0.023 $\pm$ 0.005
	Infeasible ( $\downarrow$ )	3000	0.141 $\pm$ 0.166	0.0001 $\pm$ 0.0004
	Feasible ( $\uparrow$ )	100	0.152 $\pm$ 0.031	0.301 $\pm$ 0.068
Multi-	Feasible ( $\uparrow$ )	3000	0.721 $\pm$ 0.021	0.298 $\pm$ 0.030

TABLE II : Mean evaluation values for feasible and infeasible demonstration data across training iterations, assigned using EF-GAIfO and WGAIfO trained with 20 different random seeds in the point-mass experiments. The evaluation values should be high for feasible demonstrations and low for infeasible demonstrations.

Panel of Figure 6 — Fig. 4 : Feasibility weights estimated by EF-GAIfO in the point-mass experiments. (a) and (b) show the multi-speed setting at 100 and 3000 training iterations, respectively, while (c) and (d) show the multi-path setting. The color of each demonstration trajectory represents its estimated feasibility weight.

Fig. 5 : Object-reaching-and-picking task setting in the quadruped robot environment.

Panel of Figure 8 — Fig. 6 : Demonstration for robot experiment

Panel of Figure 9 — Fig. 7 : Human-collected object-reaching-and-picking demonstrations used in the quadruped robot experiment. The demonstrations include variations in planar position and pitch motion.

	EF-GAIfO	WGAIfO	GAIfO
Target 1	$\mathbf{83.6\pm 9.2\%}$	$34.8\pm 47.7\%$	$37.4\pm 14.2\%$
Target 2	$\mathbf{88.0\pm 5.7\%}$	$35.2\pm 47.8\%$	$36.8\pm 8.1\%$
Target 3	$\mathbf{84.8\pm 10.0\%}$	$39.0\pm 48.8\%$	$42.0\pm 17.6\%$
Random target	$\mathbf{85.8\pm 7.4\%}$	$26.4\pm 44.1\%$	$35.4\pm 18.1\%$

TABLE III : Object-reaching-and-picking task success rates in simulation for three fixed target positions and a random target setting. Each value is the mean success rate over five policies trained with different random seeds, with each policy evaluated over 100 trials.

Panel of Figure 11 — Fig. 8 : Execution of the learned policies in the object-reaching-and-picking task using the Go2 robot in simulation. (a) EF-GAIfO successfully completes the task. (b) WGAIfO fails due to excessive pitch motion. (c) GAIfO fails due to misalignment between the gripper and the object.

Panel of Figure 12 — Fig. 9 : Feasibility weights estimated by EF-GAIfO for demonstrated pitch motions in the quadruped robot experiment at (a) 100 and (b) 3000 training iterations.

	EF-GAIfO	WGAIfO	GAIfO
Target 1	70%	10%	10%
Target 2	70%	10%	10%
Target 3	80%	10%	20%
Random target	80%	30%	20%

TABLE IV : Real-world object-reaching-and-picking task success rates for three fixed target positions and a random target setting. Each setting is evaluated over 10 trials.

Panel of Figure 14 — Fig. 10 : Representative real-world execution examples of the learned policies in the object-reaching-and-picking task using the Go2 robot. The snapshots illustrate typical behaviors observed for each method: (a) EF-GAIfO successfully approaches, grasps, and lifts the object; (b) WGAIfO exhibits a grasping failure; and (c) GAIfO exhibits a grasping failure caused by gripper–object misalignment.

Explore similar work

Learning Beyond What Humans Can Demonstrate
Sep 21, 2026Yuchen Song, Aditya Mittal, Unnat JainTeleoperation Behavior Cloning
University of California, Irvine
Demo-JEPA: Joint-Embedding Predictive Architecture for One-shot Cross-Embodiment Imitation
May 20, 2026Jingyang He, Guangrun Li, Jieyu Zhang +3Cross-Embodiment Transfer Embodiment
State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University · Beijing Innovation Center of Humanoid Robotics · University of Washington
Good in Bad (GiB): Sifting Through End-user Demonstrations for Learning a Better Policy
May 2, 2026Noushad Sojib, Ola Ghattas, Momotaz BegumImitation Learning Bad
Department of Computer Science, University of New Hampshire, USA

Sep 21, 2026, cs.ROJ/K move · Enter open · S save

Learning Beyond What Humans Can Demonstrate

Yuchen Song, Aditya Mittal, Unnat Jain

University of California, Irvine

Teleoperation Behavior Cloning Guardrail Latent Failure Patterns

Open paper AskPDF arXiv

Behavior cloning for robot manipulation relies on expert demonstrations. However, for tasks that require dynamic stability, precise contact timing, or dexterous coordination, human operators may find it hard or even impossible to collect data. We study this infeasible-demonstration regime and propose GLIDE: Guardrails for Learning from Infeasible Demonstrations Efficiently, a framework that infers task-specific failure modes and converts them into executable guardrails for data collection and policy deployment. Given a task description and the conditioning teleoperation code, GLIDE writes guardrails that use system states to filter teleoperation and policy commands, constrain failure-prone actions, and iteratively improve from trajectory feedback. Across three tasks, GLIDE discovers emergent guardrails that go beyond domain-expert hardcoded ones, improving data collection over naive VR teleoperation and domain-expert hardcoded guardrails. After refinement, GLIDE raises data-collection success from 0-10 percent to 70-90 percent across the three tasks. During policy execution, mixed-data guarded policies reach 70 percent, 60 percent, and 60 percent success on Tomato plate transfer, Marker handover and stand, and Wine serving tasks. These results show that GLIDE can support policy learning when direct demonstrations are infeasible. Project website: http://guardrail-policy.github.io/

Abstract

Explore similar work

Learning Beyond What Humans Can Demonstrate

Demo-JEPA: Joint-Embedding Predictive Architecture for One-shot Cross-Embodiment Imitation

Good in Bad (GiB): Sifting Through End-user Demonstrations for Learning a Better Policy