Hanbit Oh*†, Masaki Murooka*,
Tomohiro Motoda, Ryoichi Nakajo, and Yukiyasu Domae
*: Equal contribution; †: Corresponding author
Under review
Imitation learning is a promising paradigm for training robot agents; however, standard approaches typically require substantial data acquisition---via numerous demonstrations or random exploration---to ensure reliable performance. Although exploration reduces human effort, it lacks safety guarantees and often results in frequent collisions---particularly in clearance-limited tasks (e.g., peg-in-hole)---thereby, necessitating manual environmental resets and imposing additional human burden. This study proposes Self-Augmented Robot Trajectory (SART), a framework that enables policy learning from a single human demonstration, while safely expanding the dataset through autonomous augmentation. SART consists of two stages: (1) human teaching only once, where a single demonstration is provided and precision boundaries---represented as spheres around key waypoints---are annotated, followed by one environment reset; (2) robot self-augmentation, where the robot generates diverse, collision-free trajectories within these boundaries and reconnects to the original demonstration. This design improves the data collection efficiency by minimizing human effort while ensuring safety. Extensive evaluations in simulation and real-world manipulation tasks show that SART achieves substantially higher success rates than policies trained solely on human-collected demonstrations.
Fig1. Overview of SART: A human provides a single demonstration and explicitly annotates precision boundaries around key waypoints. The robot then autonomously augments its dataset by exploring within these boundaries, generating diverse, collision-free trajectories used to learn a robust policy.
Our key contributions are as follows:
Self-augmented robot trajectory (SART) is introduced, which is a framework that enables robots to safely augment their own training data within human-annotated spatial boundaries derived from a single demonstration.
By autonomously generating diverse trajectories within these safe boundaries, SART efficiently expands the demonstration dataset and yields more robust policies than those trained on human demonstrations alone.
Extensive evaluations in both simulation and physical robot experiments show that SART improves learning efficiency and achieves higher success rates than fully human-supervised baselines in clearance-limited tasks (e.g., peg-in-hole, door opening, and lid opening).
Evaluation of the SART framework, as indicated in Fig. 2, was conducted in both simulated and real-world environments. The environment and task definitions are as follows:
Fig.2: Environments and Tasks
Single-demo Replay: Direct replay of a single teleoperated trajectory without any additional learning or augmentation.
BC [bojarski et al.]: Standard IL from multiple human demonstrations, referred to as BC. The demonstrations were collected under randomized goal-object positions in accordance with each task setting.
Contact-free MILES [papagiannis et al.]: A self-augmented IL method from a single demonstration, based on MILES. Although MILES is closely related to SART, it differs in two critical ways—(i) the augmentation boundary is fixed, regardless of task geometry, with a sphere of 2cm radius placed at every demonstration waypoint rather than selectively; and (ii) augmented trajectories are generated by returning to the center of each sphere, instead of progressing toward the convergence waypoint. In addition, unlike the original MILES formulation, which assumes access to costly contact-sensing systems for handling contact-rich scenarios, the comparison here is conducted under a lower-cost setup without contact sensing.
A visual inspection of the demonstrations and rollouts revealed fundamental differences across the baselines, as follows:
Fig.3: Comparison of training data (N=30) for the peg-in-hole task across varying methods
Fig.4: Rollouts of policies trained using N = 30 training data across various methods for the peg-in-hole task
Single-demo Replay
BC
Contact-free MILES
SART (proposed)
Single-demo Replay
BC
Contact-free MILES
SART (proposed)
Single-demo Replay
BC
Contact-free MILES
SART (proposed)
Single-demo Replay
BC
Contact-free MILES
SART (proposed)
Single-demo Replay
BC
Contact-free MILES
SART (proposed)
Single-demo Replay
BC
Contact-free MILES
SART (proposed)
The SART framework was evaluated in both simulated and real environments with a focus on clearance-limited robotic tasks. The following three primary questions guided the analysis:
Table 1. Quantitative comparison of SART against baselines on clearance-limited manipulation tasks.
Fig5. Comparison of success rates with respect to the number of training data (N ) for SART and baseline methods on clearance-limited robotic manipulation tasks.
Single Demo Replay: Fails under randomized object positions, unable to adapt beyond one trajectory.
BC: Uses multiple human demos and generalizes somewhat, but compounding errors limit success to ~50–60%.
Contact-Free MILES: Augments a single demo with fixed spheres but adds backward motions and poor coverage, leading to stalls and near-zero success as data grows.
SART: From one human demo + annotated boundaries, generates diverse forward-directed trajectories that consistently succeed, achieving top success in both simulation and real-world tasks, nearly doubling BC with far less human effort.
Fig6. Comparison of SART with ablated variants. Only simulation results are reported.
Takeaway: SART’s strong performance comes from combining position augmentation, convergence-based trajectory design, and merging with the demonstration. Orientation augmentation is less critical in our current benchmarks but may become central in tasks where orientation diversity is important.
Fig7. Comparison of average duration for dataset collection in both simulation and real-world settings.
Simulation: SART cuts average collection time from ~30.5 min (BC) to ~17.8 min, despite no reset overhead.
Real world: Reduces duration from ~11.5 min (BC) to ~5.4 min, even with resets.
Takeaway: SART substantially lowers both human involvement and overall dataset generation time, while still achieving higher success rates.
