Self-Augmented Robot Trajectory (SART)

• Self-Augmented Robot Trajectory (SART) is a family of methods that augment robot trajectories using self-supervised models, online timing adaptations, and demonstration-based safe augmentation.
• These techniques exploit additional structure beyond direct imitation to rectify predictions, adjust motion timing with human interaction, or safely expand demonstration datasets.
• SART approaches improve trajectory planning performance and efficiency, yielding measurable gains in success rates, runtime consistency, and reduced data collection times.

Self-Augmented Robot Trajectory (SART) denotes a family of trajectory-generation and trajectory-adaptation schemes in which the robot augments its own effective supervision or execution by exploiting structure that is not provided as dense external labels. In the cited literature, the term covers at least three closely related formulations: a self-supervised recurrent trajectory model that uses paired forward and inverse kinematics to generate rectified training targets (Cibula et al., 2 Jul 2025), a trajectory-advancement mechanism in physical human-robot collaboration that reparameterizes motion timing online using a helpful interaction wrench (Tirupachuri et al., 2019), and a safe imitation-learning framework that expands a single human demonstration through demonstrator-annotated precision boundaries and autonomous collision-free augmentation (Oh et al., 11 Sep 2025). Across these formulations, the common element is self-augmentation: the trajectory is modified, extended, or supervised by robot-internal simulation, bounded interaction-aware timing adjustment, or robot-generated data rather than by dense externally supplied trajectory labels.

1. Conceptual scope and defining principle

SART is not presented as a single monolithic algorithm across the literature. Rather, it is a unifying principle under which a robot augments trajectory generation or trajectory execution using additional structure beyond direct imitation. In the self-supervised recurrent formulation, the robot augments training targets through paired Forward Model (FM) and Inverse Model (IM) modules that internally rectify predicted waypoints (Cibula et al., 2 Jul 2025). In physical human-robot collaboration, the robot augments trajectory timing by detecting a helpful component of the human interaction wrench and advancing a path progression variable $s$ faster than nominal (Tirupachuri et al., 2019). In safe imitation learning, the robot augments a single human demonstration by generating diverse, collision-free trajectories inside demonstrator-annotated precision spheres and reconnecting them to the original demonstration (Oh et al., 11 Sep 2025).

A concise way to characterize SART is therefore as trajectory self-augmentation under constraints. The augmentation may occur in supervision space, as when internal models produce a rectified target sequence; in execution time, as when a helpful wrench accelerates motion along a reference path; or in dataset space, as when safe robot-generated rollouts expand a demonstration set. This suggests that SART is best understood as a design pattern rather than a single architecture.

The three formulations also differ in what is held fixed. The self-supervised recurrent method fixes endpoint supervision and learns intermediate structure (Cibula et al., 2 Jul 2025). The pHRC formulation keeps the spatial path $x_d(s)$ intact while altering the timing law $s(t)$ (Tirupachuri et al., 2019). The imitation-learning formulation preserves a human demonstration and its task structure while varying locally around selected key waypoints inside annotated spheres (Oh et al., 11 Sep 2025). In each case, augmentation is constrained rather than unconstrained: by FM/IM consistency, by bounded advancement and selective wrench cancellation, or by human-validated geometric boundaries.

2. Self-supervised recurrent SART via forward and inverse models

In "Towards Bio-Inspired Robotic Trajectory Planning via Self-Supervised RNN" (Cibula et al., 2 Jul 2025), the trajectory model (TM) is trained without ground-truth intermediate waypoints. The state is defined as

$s(t) = [\theta(t), ef(t)],$

where $\theta(t)$ is the joint configuration and $ef(t)$ is the end-effector position. Actions are joint increments,

$a(t) = \theta(t+1) - \theta(t).$

The kinematic modules are expressed conceptually as

$x = f(q), \qquad q = g(x).$

The TM maps initial and goal states to a predicted intermediate sequence,

$$TM: [s(0), s(T)] \to \hat{\tau}_s = [\hat{s}(1), \ldots, \hat{s}(T-1)].$$

For each predicted $\hat{s}(t)$, the IM computes an action from the last rectified state $x_d(s)$0 to the predicted waypoint,

$x_d(s)$1

and the FM then simulates the outcome,

$x_d(s)$2

The rectified trajectory $x_d(s)$3 becomes the supervisory signal for the TM. The paper describes this as a cognitively inspired self-supervised learning scheme, and the synthesized connection explicitly identifies it as a concrete instantiation of SART (Cibula et al., 2 Jul 2025).

The loss used for TM training compares predicted waypoints to the rectified trajectory and anchors the ends: $x_d(s)$4

$x_d(s)$5

The details further note a conceptual link to cycle-consistency. While the implementation does not explicitly optimize

$x_d(s)$6

the IM$x_d(s)$7FM rectification loop enforces practical consistency by penalizing predictions whose simulated execution does not land at a consistent rectified state (Cibula et al., 2 Jul 2025).

The architecture is a GRU decoder with $x_d(s)$8, $x_d(s)$9, and $s(t)$0 activations, followed by a time-distributed prediction head. The TM predicts only end-effector positions $s(t)$1; joint configurations $s(t)$2 are handled via FM and IM during rectification and later for action generation. Training uses 12,000 endpoint pairs from KUKA LBR iiwa, 7 DoF trajectories in myGym, with sequence length $s(t)$3 and no ground-truth intermediates supplied to the TM (Cibula et al., 2 Jul 2025).

The reported FM and IM training characteristics are specific: FM was trained with Adam, $s(t)$4, for 60 epochs, reaching $s(t)$5 mm and $s(t)$6 rad; IM used AdamW, $s(t)$7, weight decay $s(t)$8, for 100 epochs, reaching $s(t)$9 rad. For TM training, 50 epochs were run over 12,000 endpoint pairs, with Adam, RMSprop, and SGD tested; the best-performing configuration was RMSprop with $s(t) = [\theta(t), ef(t)],$0 (Config #4), with final $s(t) = [\theta(t), ef(t)],$1 (Cibula et al., 2 Jul 2025).

This formulation is explicitly bio-inspired. The paired FM/IM structure is linked to motor control theories associated with Wolpert and Kawato, the method is described as consistent with the uncontrolled manifold concept because TM focuses on task-relevant variables $s(t) = [\theta(t), ef(t)],$2, and the overall progression from motor babbling for FM/IM to higher-level trajectory generation is framed as developmental robotics (Cibula et al., 2 Jul 2025).

3. Trajectory advancement in physical human-robot collaboration

In "Trajectory Advancement during Human-Robot Collaboration" (Tirupachuri et al., 2019), SART appears as online trajectory self-augmentation in execution rather than learning. The robot tracks a spatial reference path $s(t) = [\theta(t), ef(t)],$3 in task space, parameterized by a scalar progression variable $s(t) = [\theta(t), ef(t)],$4. Under nominal timing, $s(t) = [\theta(t), ef(t)],$5, giving a conventional time-parameterized reference $s(t) = [\theta(t), ef(t)],$6. The central idea is to exploit a helpful human interaction wrench so that the robot advances along the path faster than nominal, completing the task sooner while preserving tracking stability considerations (Tirupachuri et al., 2019).

The robot is modeled as a free-floating multi-body system with dynamics

$s(t) = [\theta(t), ef(t)],$7

For the controlled end-effector,

$s(t) = [\theta(t), ef(t)],$8

The helpfulness test is based on the interaction wrench $s(t) = [\theta(t), ef(t)],$9 projected onto the desired motion direction. Let

$\theta(t)$0

and define

$\theta(t)$1

Equivalently, helpfulness can be assessed through task-space power

$\theta(t)$2

Interaction is helpful when $\theta(t)$3, hindering when $\theta(t)$4, and agnostic when $\theta(t)$5 (Tirupachuri et al., 2019).

The nominal desired task acceleration is

$\theta(t)$6

with $\theta(t)$7 and $\theta(t)$8. A compact feedback-linearizing torque law is reported as

$\theta(t)$9

In its nominal form, this law cancels the acceleration effect of the external wrench completely, which is safe but precludes active exploitation of human assistance (Tirupachuri et al., 2019).

To incorporate helpful interaction, the paper defines a correction wrench

$ef(t)$0

and augments the desired task acceleration by a tangential component: $ef(t)$1 The torque law then compensates only the non-helpful part of the external wrench: $ef(t)$2 In parallel, the path parameter is advanced faster than nominal through a bounded law. The paper gives

$ef(t)$3

with $ef(t)$4 used in practice (Tirupachuri et al., 2019).

A Lyapunov-style argument yields the sufficient condition

$ef(t)$5

for safe advancement. The details emphasize that a formal proof is not given because the modified law introduces an algebraic loop involving $ef(t)$6, $ef(t)$7, and $ef(t)$8; implementation resolves this through delayed and filtered signals (Tirupachuri et al., 2019). This is a notable limitation of the formulation: its stability case is numerically supported rather than formally closed.

Experimental validation used the iCub humanoid, with the right foot tracking a 1D Cartesian minimum-jerk sinusoid along $ef(t)$9, amplitude $a(t) = \theta(t+1) - \theta(t).$0 m and frequency $a(t) = \theta(t+1) - \theta(t).$1 Hz, with the base fixed. In both Gazebo simulation and hardware experiments, assistive wrenches caused $a(t) = \theta(t+1) - \theta(t).$2 to increase and $a(t) = \theta(t+1) - \theta(t).$3 to rise above nominal, thereby advancing $a(t) = \theta(t+1) - \theta(t).$4. Agnostic wrenches did not alter $a(t) = \theta(t+1) - \theta(t).$5 or $a(t) = \theta(t+1) - \theta(t).$6. On the real robot, when the sinusoid reversed direction, a sustained $a(t) = \theta(t+1) - \theta(t).$7 wrench became agnostic and advancement stopped, demonstrating directional gating. Tracking error increased relative to simulation because of friction and phase delays, but amplitudes were reached (Tirupachuri et al., 2019).

This formulation shows that SART need not imply learned trajectory synthesis. It can also denote online trajectory time-scaling that selectively refrains from canceling helpful external forces.

4. Safe self-augmentation for imitation learning from a single demonstration

In "Self-Augmented Robot Trajectory: Efficient Imitation Learning via Safe Self-augmentation with Demonstrator-annotated Precision" (Oh et al., 11 Sep 2025), SART is a two-stage imitation-learning framework that learns from a single human demonstration while safely expanding the dataset through autonomous augmentation. The motivating problem is that behavioral cloning from few demonstrations suffers from covariate shift, whereas exploration-based augmentation is risky in clearance-limited tasks such as peg-in-hole, door or lid opening, and tight placing, often causing collisions and repeated manual resets (Oh et al., 11 Sep 2025).

The demonstration is recorded as either a sequence $a(t) = \theta(t+1) - \theta(t).$8 or, in the implementation, as a dataset

$a(t) = \theta(t+1) - \theta(t).$9

Observations $x = f(q), \qquad q = g(x).$0 are RGB images from a wrist-mounted camera plus gripper state. Actions $x = f(q), \qquad q = g(x).$1 are desired 6-DoF end-effector motion and gripper state, expressed relative to the end-effector pose at $x = f(q), \qquad q = g(x).$2 (Oh et al., 11 Sep 2025).

The key annotation mechanism consists of selected key waypoints and precision boundaries represented as spheres: $x = f(q), \qquad q = g(x).$3 In the implementation, $x = f(q), \qquad q = g(x).$4 is the end-effector position $x = f(q), \qquad q = g(x).$5 and $x = f(q), \qquad q = g(x).$6. The demonstrator steps through the demonstration’s 3D point cloud, derived from the wrist RGB-D camera, together with robot URDF visualization, and adjusts $x = f(q), \qquad q = g(x).$7 so that spheres permit allowable variation while avoiding the environment geometry. Small $x = f(q), \qquad q = g(x).$8 are used near contact-sensitive regions, and larger $x = f(q), \qquad q = g(x).$9 in free space (Oh et al., 11 Sep 2025).

Self-augmentation samples target poses on these spheres. Position is sampled uniformly on $$TM: [s(0), s(T)] \to \hat{\tau}_s = [\hat{s}(1), \ldots, \hat{s}(T-1)].$$0 by drawing

$$TM: [s(0), s(T)] \to \hat{\tau}_s = [\hat{s}(1), \ldots, \hat{s}(T-1)].$$1

then setting

$$TM: [s(0), s(T)] \to \hat{\tau}_s = [\hat{s}(1), \ldots, \hat{s}(T-1)].$$2

Orientation is perturbed by a bounded random rotation with $$TM: [s(0), s(T)] \to \hat{\tau}_s = [\hat{s}(1), \ldots, \hat{s}(T-1)].$$3. The convergence waypoint is defined as the first waypoint after the selected key pose that lies outside the sphere: $$TM: [s(0), s(T)] \to \hat{\tau}_s = [\hat{s}(1), \ldots, \hat{s}(T-1)].$$4 The convergence pose is $$TM: [s(0), s(T)] \to \hat{\tau}_s = [\hat{s}(1), \ldots, \hat{s}(T-1)].$$5 (Oh et al., 11 Sep 2025).

Augmented trajectories are generated by executing a straight-line motion from $$TM: [s(0), s(T)] \to \hat{\tau}_s = [\hat{s}(1), \ldots, \hat{s}(T-1)].$$6 to a sampled pose $$TM: [s(0), s(T)] \to \hat{\tau}_s = [\hat{s}(1), \ldots, \hat{s}(T-1)].$$7 over a fixed duration $$TM: [s(0), s(T)] \to \hat{\tau}_s = [\hat{s}(1), \ldots, \hat{s}(T-1)].$$8 s, recording the sequence at 10 Hz, and reversing the segment so that it leads from $$TM: [s(0), s(T)] \to \hat{\tau}_s = [\hat{s}(1), \ldots, \hat{s}(T-1)].$$9 back to $\hat{s}(t)$0. The final trajectory is

$\hat{s}(t)$1

The details state that reconnecting at $\hat{s}(t)$2 matches both end-effector pose and joint configuration of the original demonstration, improving position/orientation and velocity continuity at the join and stabilizing subsequent execution (Oh et al., 11 Sep 2025).

Policy learning uses behavioral cloning with squared action error: $\hat{s}(t)$3 The policy architecture comprises a fine-tuned ResNet18 visual encoder with image feature dimension 512, gripper feature concatenation, and an MLP head with 2 hidden layers of 512 units each, ReLU activation, AdamW optimizer, learning rate $\hat{s}(t)$4, weight decay $\hat{s}(t)$5, minibatch size 32, and 40 epochs. Observation history length is 2 (Oh et al., 11 Sep 2025).

The empirical results are reported on MuJoCo simulation with UR5e + Robotiq 2F-85 across four clearance-limited tasks and on real-world UR5e + Robotiq Hand-E for bottle placing and lid closing. Success rates averaged over 100 simulation runs per method-task and 30 real runs per task are summarized below.

Method	Simulation / real success rates	Average
Single-demo replay	peg 4%, door 6%, lid 28%, toolbox 16%; real bottle 0%, lid close 30%	$\hat{s}(t)$6
Behavioral cloning (multi-demo)	peg 28%, door 58%, lid 31%, toolbox 8%; real bottle 67%, lid close 60%	$\hat{s}(t)$7
Contact-free MILES	peg 41%, door 0%, lid 0%, toolbox 2%; real bottle 53%, lid close 43%	$\hat{s}(t)$8
SART	peg 89%, door 88%, lid 83%, toolbox 67%; real bottle 80%, lid close 93%	$\hat{s}(t)$9

The paper reports that SART significantly outperforms all baselines across tasks ($x_d(s)$00-test, $x_d(s)$01); in simulation its mean success is more than $x_d(s)$02 behavioral cloning, and in real tasks it reaches $x_d(s)$03–$x_d(s)$04 versus behavioral cloning’s $x_d(s)$05–$x_d(s)$06 (Oh et al., 11 Sep 2025). The method also reduces data-collection duration: in simulation, SART requires 17.77 min versus 30.50 min for behavioral cloning; in real experiments, 5.40 min versus 11.50 min. It requires only one environment reset after the single demo (Oh et al., 11 Sep 2025).

5. Common mechanisms, computational properties, and empirical behavior

Although the three formulations differ substantially, they share an operational motif: augmentation is driven by structured constraints rather than unrestricted search. In the self-supervised recurrent case, augmentation arises from internal simulation with frozen FM and IM modules trained by motor babbling (Cibula et al., 2 Jul 2025). In the pHRC case, augmentation is driven by a helpful component of measured interaction wrench projected onto the current trajectory tangent, with $x_d(s)$07 bounded below by 1 and above by $x_d(s)$08 (Tirupachuri et al., 2019). In the imitation-learning case, augmentation is geometrically constrained to human-validated spheres and reconnects to the demonstration through convergence waypoints (Oh et al., 11 Sep 2025).

The bounded-computation property is explicit in the recurrent planner. The GRU decoder produces a fixed-length trajectory in a single forward pass, with one rectification pass via IM$x_d(s)$09FM per waypoint during training. At inference, only the decoder is required unless online correction is added, yielding predictable runtime; empirical inference time for a 10-step trajectory is reported as approximately $x_d(s)$10 on an RTX 4080 SUPER (Cibula et al., 2 Jul 2025). This is contrasted with PRM/RRT planning times, which may be unpredictable and can scale poorly with constraints or environment complexity (Cibula et al., 2 Jul 2025).

The reported empirical quality of the recurrent planner is mixed in a way characteristic of proof-of-concept studies. For Config #4, endpoint adherence was approximately $x_d(s)$11 m $x_d(s)$12 m at the initial end and $x_d(s)$13 m $x_d(s)$14 m at the final end; average spacing was approximately $x_d(s)$15 m $x_d(s)$16 m; maximum spacing deviation was approximately $x_d(s)$17 m $x_d(s)$18 m; and smoothness, measured as average angle, was approximately $x_d(s)$19 (Cibula et al., 2 Jul 2025). These values indicate that the model can produce curved trajectories beyond linear interpolation, but endpoint adherence and spacing regularity remain imperfect.

In the pHRC setting, the principal behavioral result is qualitative rather than statistical. Assistive wrenches increase $x_d(s)$20 and advance $x_d(s)$21 beyond the nominal $x_d(s)$22; agnostic wrenches do not. The paper further notes a magnitude correlation: larger assistive wrench yields larger $x_d(s)$23 and longer advancement (Tirupachuri et al., 2019). Statistical significance is not reported; repeatability is demonstrated across trials.

In safe imitation learning, the scaling behavior with dataset size is also reported. As $x_d(s)$24 increases in simulation, behavioral cloning plateaus below approximately $x_d(s)$25 success, contact-free MILES degrades toward near-zero due to backward motions, and SART remains superior for $x_d(s)$26 (Oh et al., 11 Sep 2025). This indicates that augmentation quality, not merely augmentation quantity, is central to the framework.

6. Limitations, misconceptions, and research directions

A common misconception would be to treat SART as synonymous with a single technique. The literature instead supports at least three distinct meanings: self-supervised trajectory learning via FM/IM rectification, human-assisted trajectory advancement by time-scaling, and safe dataset augmentation for behavioral cloning (Cibula et al., 2 Jul 2025, Tirupachuri et al., 2019, Oh et al., 11 Sep 2025). What unifies them is the self-augmentation principle, not a fixed algorithmic stack.

The recurrent self-supervised planner has several explicit limitations. It is a proof-of-concept with no explicit modeling of obstacles, collisions, or joint limits; trajectory length is fixed; endpoint adherence is imperfect; spacing irregularity and sharp turns occur in many trajectories; and the TM predicts only $x_d(s)$27, not $x_d(s)$28, in the trajectory head (Cibula et al., 2 Jul 2025). The details suggest several SART-style extensions, including explicit

$x_d(s)$29

joint-limit penalties

$x_d(s)$30

collision penalties

$x_d(s)$31

and Jacobian-based shaping through

$x_d(s)$32

These terms were not used in the reported experiment (Cibula et al., 2 Jul 2025). A plausible implication is that the self-supervised SART formulation remains most credible, in its present form, for unconstrained kinematic planning rather than cluttered environments.

The pHRC advancement method likewise has explicit limitations. The path must remain feasible under faster timing; accurate Jacobians and dynamic terms are assumed; the interaction is measured at the end-effector; and the method has an algebraic loop without a formal proof of stability. Sensitivity to frame misalignment, friction, and delay is noted, and tuning of $x_d(s)$33, $x_d(s)$34, filtering, and thresholds is necessary (Tirupachuri et al., 2019). The details identify future directions including rigorous passivity or stability proofs, adaptive thresholds and gains, multi-contact scenarios, and integration with impedance or admittance control (Tirupachuri et al., 2019).

The safe imitation-learning formulation also makes its limits explicit. Formal collision-free guarantees are not provided; safety is empirical, based on human-validated spheres and conservative augmentation; the method does not explicitly model contact forces; and it is described as effective for clearance-limited rather than strongly contact-rich tasks (Oh et al., 11 Sep 2025). Its performance depends on correct human annotation of spheres and appropriate waypoint density. Ablation results show significant degradation without position augmentation, with return-to-center reconnection instead of convergence waypoint reconnection, and without merging the demonstration suffix; removing orientation augmentation causes a smaller degradation (Oh et al., 11 Sep 2025). This indicates that SART’s gains arise from a specific combination of local safe diversity and preserved downstream task progress.

Taken together, the literature suggests three broad research directions. First, incorporate explicit feasibility structure—joint limits, collisions, task dynamics, and contact models—into internal self-augmentation loops (Cibula et al., 2 Jul 2025). Second, formalize stability and passivity for interaction-driven advancement schemes that currently rely on filtering and delayed implementation (Tirupachuri et al., 2019). Third, automate or partially automate safe augmentation boundary specification while preserving the empirical safety advantages of demonstrator-annotated precision (Oh et al., 11 Sep 2025). These directions are consistent with the view that SART is a general strategy for bounded, structured trajectory augmentation rather than a closed research problem.