AdapTraj: Adaptive Trajectory Frameworks

• AdapTraj is an adaptive trajectory framework that integrates methods like LLM-driven code generation, latent visual prompts, and causal decomposition for robust trajectory modifications.
• It supports diverse applications from robotic manipulation and human motion prediction to multi-agent forecasting and real-time robot control in dynamic environments.
• Empirical evaluations demonstrate significant improvements in constraint satisfaction, reduced ADE, and safe, real-time adaptations across challenging scenarios.

AdapTraj encapsulates several distinct methodologies for trajectory adaptation, prediction, and uncertainty quantification across robotics, human motion modeling, and time-series forecasting. Despite the diversity of technical mechanisms, the unifying theme is the deployment of adaptive modules—ranging from language-driven code generation and prompt-based scene conditioning to causal domain decomposition—to robustly tailor or predict trajectories in dynamically changing or previously unseen environments. This entry details the principal AdapTraj frameworks, their technical foundations, and empirical outcomes as evidenced by key arXiv sources.

1. Language-Driven Trajectory Adaptation via LLMs

AdapTraj, as introduced in "Trajectory Adaptation using LLMs," is a two-stage LLM-based pipeline for flexible robotic trajectory adaptation (Maurya et al., 17 Apr 2025). The system translates natural language instructions into concrete manipulations of generic motion trajectories, such as those from RRT, A* or human demonstrations, with no task-specific training required.

• Data Flow: The framework operates on input trajectory $T = \{w_1, ..., w_N\}$ where $w_i = (x_i, y_i, z_i, v_i)$, environment object set $O = \{(\mathrm{Label}(O_j), P(O_j))\}$, and a free-form instruction $L_{\mathrm{instruct}}$.
• Architecture:

1. High-Level Plan (HLP) Generation: LLM is prompted with function/coordinate schema, in-context examples, and $L_{\mathrm{instruct}}$ to synthesize a JSON plan that specifies geometric or kinematic edits.
2. Code Generation: A subsequent LLM invocation uses the plan to generate Python code that modifies waypoints and produces the adapted trajectory $T_{mod}$, enforcing constraints (start/goal preservation, spatial buffer, smoothness, velocity continuity) entirely through procedural logic—no model re-training.

• Interpretability: The approach promotes transparency by exposing both HLP and code for human review and iterative feedback, facilitating error diagnosis or rapid refinement.
• Empirical Performance: On ^LaTTe^ instruction sets and compound commands, the method significantly exceeds feature-based seq2seq models, e.g., achieving 92% vs. 75% constraint satisfaction and lower smoothness error (0.045 m vs. 0.087 m SE).

2. Scene-Adaptive Human Trajectory Prediction via Latent Corridors

In "Adaptive Human Trajectory Prediction via Latent Corridors," AdapTraj describes a lightweight, few-shot adaptation mechanism for human trajectory predictors, critical where test-time scene factors differ from pre-training (Thakkar et al., 2023).

• Mechanism: A scene-specific latent prompt $p \in \mathbb{R}^{h \times w}$ (the "latent corridor") is learned and summed into the history heatmaps input to the (frozen) base predictor, enabling the system to encode and exploit local spatiotemporal peculiarities (e.g., crowd surges, temporary obstacles).
• Parameterization: $p$ is stored as a rank-1 outer product $u v^\top$, introducing only $\sim$0.1% additional parameters per scene.
• Learning: On-scene adaptation is performed by optimizing only $(u, v)$ (and optionally the final predictor head), using pixelwise BCE loss on a small deployment sample—typically under five minutes or a few hundred person-seconds.
• Results: On MOTSynth, MOT, and WildTrack, the method yields up to 24% ADE improvement and qualitative gains in compliance with walkable surfaces, navigational affordances, and time-varying crowd behaviors.

3. Multi-Source Domain Generalization for Multi-Agent Trajectory Prediction

AdapTraj in "A Multi-Source Domain Generalization Framework for Multi-Agent Trajectory Prediction" targets robust trajectory prediction under domain shift (Qian et al., 2023). The framework is grounded in an explicit causal decomposition and modular architecture:

• Structural Decomposition: Each agent/history is split into domain-invariant $(H^i)$ and domain-specific $(H^s)$ features, for both the focal and neighboring agents. These are extracted via parallel MLP/Transformer modules, with a mixture-of-experts formulation for domain-specificity.
• Integration: Both sets of features are softly enforced to be orthogonal and domain-invariant components are adversarially aligned to the domain index.
• Optimization: The composite loss includes base prediction, reconstruction, adversarial domain classification, and orthogonality penalties.
• Efficacy: On ETH/UCY, L-CAS, SYI, and SDD datasets, AdapTraj consistently outperforms vanilla and other domain generalization baselines, improving ADE and FDE with negligible inference-time cost.

4. Adaptive Trajectory Sampling for Socially Aware Robot Person Following

In "Adap-RPF: Adaptive Trajectory Sampling for Robot Person Following in Dynamic Crowded Environments," AdapTraj constitutes the trajectory sampling and evaluation core (Situ et al., 13 Oct 2025).

• Sampling: A Sobol sequence generator populates a semi-annular, socially compliant region (defined by proxemics) behind or beside the target; points are filtered for occupancy and scored by a weighted sum of occlusion, proximity, nominal distance, travel, and stickiness costs.
• Optimization: The lowest-cost point defines the robot's follow offset, which is tracked against the predicted target trajectory using a Model Predictive Path Integral (MPPI) controller with proactive avoidance of predicted pedestrian locations.
• Outcomes: Across multiple crowd densities, AdapTraj outperforms fixed and sparse following baselines in target visibility, safety, comfort (TPZ), and robustness, while supporting real-time (>50 Hz) deployment.

5. Adaptive Conformal Prediction Intervals for Trajectory Ensembles

AdapTraj as proposed in "Adaptive Conformal Prediction Intervals Over Trajectory Ensembles" formalizes statistically valid, adaptive uncertainty quantification for trajectory ensembles in forecasting scenarios (Li et al., 18 Aug 2025).

• Calibrated Coverage: Using PCP-based non-conformity scores, sets $C^h_t$ are constructed at each forecast horizon $h$ to guarantee $1-\alpha$ coverage.
• Online Adaptivity: The target miscoverage $\alpha^h_t$ is dynamically updated per-horizon via a gradient step, and intervals $I^h_t$ are produced to account for non-stationarity. A horizon-wide optimization balances miscoverage and sharpness.
• Discontinuity: Intervals are unions of balls around ensemble members, naturally admitting disjoint (multi-modal) regions.
• Metrics: AdapTraj exhibits best or near-best sharpness and calibration scores on real-world and synthetic trajectory datasets (Lane, Cyclone, FluSight, MarkovAR), maintaining coverage under marked regime shifts.

6. Smooth Online Robot Trajectory Adaptation with Bounded Kinematics

Within "TrueÆdapt: Learning Smooth Online Trajectory Adaptation with Bounded Jerk, Acceleration and Velocity in Joint Space," the AdapTraj module focuses on physically feasible, feedback-driven adaptation in joint space (Kiemel et al., 2020).

• Constraints: At each step for every joint, acceleration outputs are clipped to respect jerk, acceleration, and velocity bounds, with analytic expressions for each constraint.
• Trajectory Generation: Clipped accelerations produce a twice-continuously differentiable trajectory segment, ensuring bounded derivatives up to jerk.
• Network: A SELU-MLP predicts per-joint accelerations based on state, sensory feedback, and future reference waypoints.
• Loss Composition: The reward combines deviation from reference, acceleration, and jerk penalties, all normalized and summed.
• execution: Runs online at 20Hz, validated on simulated and real KUKA iiwa platforms.

Comparative Overview

Approach	Domain	Core Adaptation Principle	Key Empirical Gains
LLM Trajectory Edits (Maurya et al., 17 Apr 2025)	Robot manipulation	Human-in-the-loop LLM code generation	17–36% ↑ constraint satisfaction
Latent Corridors (Thakkar et al., 2023)	Human trajectory prediction	Prompt-tuned per-scene visual correction	10–24% ↓ ADE
Multi-Source DG (Qian et al., 2023)	Multi-agent prediction	Causal decomposition, mixture-of-experts	4–15% ↓ ADE, resilient to shift
Socially Aware Sampling (Situ et al., 13 Oct 2025)	Robot person following	Proxemic, occlusion-aware dense sampling	10–40% ↑ task/visibility rates
Conformal Intervals (Li et al., 18 Aug 2025)	Forecasting, UQ	Online adaptive conformal prediction	Tight intervals, valid coverage
Physical Feasibility (Kiemel et al., 2020)	Online robot control	Bounded kinematic adaptation, neural policy	Guaranteed kinematic bounds

All variants demonstrate that adaptive approaches—whether through LLM-driven code, causal disentangling, scene-specific visual prompts, or online statistical calibration—yield substantive gains for trajectory editing, prediction, and control in non-stationary, multi-modal, or otherwise challenging environments.