ADAPT: Adaptive Systems in Robotics & AI

Updated 3 July 2026

ADAPT is a collection of adaptive systems that dynamically adjust perception, control, and representation to enhance efficiency and robustness.
It integrates frameworks from robotics, multimodal learning, and autonomous agents, achieving state-of-the-art performance in uncertain environments.
Applications span humanoid locomotion, federated learning, and adaptive scientific workflows, demonstrating practical impact across diverse domains.

ADAPT refers to a collection of modern methodologies, frameworks, and systems that leverage adaptation, adaptivity, and dynamic adjustment across a wide array of computational and physical domains. In contemporary research literature, ADAPT appears both as the name of specific algorithms and architectures and as an acronym for specialized systems, notably in robotics, multimodal learning, vision-language modeling, robust control, federated learning, and automated scientific workflows. These systems are united by the overarching objective of integrating adaptive mechanisms—across perception, control, representation, and governance—thereby enhancing robustness, efficiency, interpretability, and generalization in changing or uncertain environments.

1. ADAPT in Humanoid Locomotion and Robotics

Several ADAPT frameworks have been introduced for highly agile and robust robotic control, primarily for humanoid locomotion in unstructured 3D environments (Shao et al., 17 Mar 2026, Lyu et al., 15 Jun 2026, Harrison et al., 2017).

Adaptive Dual-Projection Architecture (ADAPT) structures real-time LiDAR point cloud observations into two low-dimensional 2D representations:
- A horizontal elevation map encodes terrain height by selecting the maximum $z$ in binned $(x, y)$ intervals within a learnable sensing radius.
- A vertical distance map records the nearest obstacles along vertical polar rays, providing information about traversable constraints.
Crucially, the spatial sensing range $r_t$ is itself a learnable action, allowing the policy to dynamically modulate its perceptual horizon—expanding in open terrain for anticipation or contracting in clutter for local detail.
Compared to high-dimensional voxel grids (≈41,000 dimensions), ADAPT’s dual-projection leads to ≈578-dimensional observations—a 71× reduction in sensor footprint—enabling threefold acceleration in training per iteration and dramatically lower GPU memory requirements.
Empirically, ADAPT achieves state-of-the-art traversal success (94.7%) on complex barriers (stairs, poles, beams) and demonstrates robust zero-shot transfer to the Unitree G1 humanoid, with minimal degradation under sensory noise or physics randomization. Performance substantially exceeds prior fixed-range and voxel-based baselines (Shao et al., 17 Mar 2026).
Analytical Disturbance-Aware Policy Training (ADAPT) incorporates an online, physically-grounded disturbance observer into whole-body humanoid RL:
- The observer estimates residual external force/torque by exploiting generalized momentum derivatives and the robot’s accessible dynamics (no force/torque sensors required).
- The disturbance signal is concatenated into the policy’s observation, and a “light-step” penalty on high lower-limb disturbance is used to encourage impact-reducing behavior.
This methodology delivers robust tracking and increased resistance to both in-distribution and out-of-distribution perturbations, demonstrated in simulation and hardware, and generalizes across unseen payloads and push scenarios (Lyu et al., 15 Jun 2026).
Earlier frameworks combined offline model-free policy learning in simulation with online tube-based model predictive control (MPC), leveraging Lipschitz-continuity guarantees to ensure both stability (invariant tubes) and bounded loss of reward during zero-shot sim-to-real policy deployment (Harrison et al., 2017).

2. ADAPT in Multimodal and Vision-Language Architectures

ADAPT is also applied to the design of models that underpin robust multimodal perception, especially in the face of domain shift and missing modalities:

AnchoreD multimodAl Physiological Transformer (ADAPT) aligns features from all modalities (e.g., biomedical signals, video, audio) in the embedding space of the best-available modality (the “anchor”) via an InfoNCE contrastive loss. A subsequent masked multimodal transformer fuses these representations, explicitly masking missing modalities at both training and inference. This approach yields linearly scalable fusion (as opposed to quadratic pairwise contrastive costs) and excels at physiological change detection even with missing sensors, outperforming traditional or imputation-based baselines (Mordacq et al., 2024).
ADAPT in Federated Prompt Learning: Adaptive Domain-Aware Prompt Tuning addresses the degradation in vision-LLM accuracy when federated over clients with disparate visual domains (e.g., different drawing styles in DomainNet) (Wei et al., 2023).
- Each client maintains both intra-domain (client-specific) and inter-domain (global) prompts, with a visual prompt module that dynamically detects domain correspondence at inference.
- Prompt fusion via domain-weighted combination enables robust, parameter-efficient alignment, with significant uplifts in multi-domain accuracy while communicating only 0.08M prompt parameters per round (vs. 86M for model finetuning).

3. ADAPT in Autonomous Agents and Commonsense Reasoning

Adaptive modules named ADAPT have been developed to equip agents with real-world commonsense and reasoning abilities:

Affordance-Based Planning (ADAPT): Explicitly benchmarks and augments embodied agents with reasoning over unsignaled, dynamic affordance constraints (Chen et al., 16 Apr 2026). The ADAPT module intercepts planner-generated subgoals, infers latent precondition satisfaction using a LoRA-finetuned multimodal vision-LLM (e.g., LLaVA-1.5B-7B), and selectively queries LLMs for corrective actions if affordances are unavailable. Empirical evaluation on the DynAfford benchmark shows marked increases in success and goal-compliance rates, especially when leveraging task-specific prompt finetuning and in-context multimodal exemplars.
As-Needed Decomposition and Planning with LLMs (ADaPT): Adopts a recursive controller that invokes a planner module only when an executor LLM fails at a specific subtask, triggering decomposition to finer subtasks (Prasad et al., 2023). This as-needed recursion dynamically adapts both to task complexity and executor capability, yielding up to 33% absolute improvements in challenging environments (ALFWorld, WebShop, TextCraft) versus non-adaptive baselines.

4. ADAPT for Efficient Computation and Simulation

Several ADAPT systems focus on optimizing computational efficiency, resource allocation, and simulation fidelity:

ADAPT as an Autoscaler: In cloud-native orchestration, ADAPT denotes a self-calibrating, closed-loop autoscaler (Baghel, 15 May 2026). It employs an online exponentially weighted moving average (EWMA) estimator of “cold start” durations (provisioning delays) to dynamically set the lookahead horizon for MPC-based scaling, enabling proactive capacity provisioning. This system—tested across six workload archetypes—outperforms both Kubernetes Horizontal Pod Autoscaler (HPA) and alternative MPC-based methods with fixed forecast horizons.
AdaPT in DNN Accelerator Emulation: AdaPT is a PyTorch-based framework providing fast, drop-in emulation and quantization-aware retraining for arbitrary approximate DNN accelerators (Danopoulos et al., 2022). Combining bitwidth-tunable quantization, support for arbitrary approximate compute units (multiply adders), and automated layer substitution, AdaPT supports up to 54× faster simulation and recovers most accuracy lost to approximation within a single training epoch, using only ≈10% of original data.

5. ADAPT in Interpretable and Multimodal Sequential Decision-Making

Several research lines employ ADAPT or closely related variants at the intersection of control and interpretability in multi-agent or multimodal sequential tasks:

Action-aware Driving cAPtion Transformer (ADAPT): An end-to-end model generating both low-level vehicle control signals and natural-language rationale for each driving action from raw video (Jin et al., 2023). Its multi-head transformer design uses shared visual representations for both captioning and control, yielding state-of-the-art results on BDD-X for automatic metrics and human understandability.
ADAPT for Trajectory Prediction: Employs an adaptive head for multi-agent, multi-future forecasting, generating dynamic, per-agent regression weights from context while maintaining a fixed parameter budget. This design attains state-of-the-art accuracy and computation (11 ms, 1.4M parameters) on major benchmarks by decoupling future endpoint regression and trajectory interpolation, substantiated via ablation studies (Aydemir et al., 2023).
Modality-Aligned Action Prompts in Navigation: In vision-language navigation, ADAPT provides explicit, CLIP-aligned action prompts (image+text pairs) and auxiliary modality-alignment and sequential-consistency objectives (Lin et al., 2022). This enables more precise stepwise phrase grounding and robust generalization, improving performance with less labeled data.

6. ADAPT in Scientific and Scholarly Processes

ADAPT also appears in frameworks addressing high-level governance, optimization, and adaptation in distributed or collaborative settings:

AI-Driven Decentralized Adaptive Publishing Testbed: Conceptualizes scholarly publishing as a feedback-controlled system, integrating authors, reviewers (human and AI), and editors through explicit, loggable policy parameters and adaptive mechanisms (Manik et al., 5 Apr 2026). The testbed adapts its governance in response to operational stressors (overload, reviewer disagreement, collusion), using closed-loop updates on triage thresholds, AI-reviewer fractions, and escalation policies while maintaining auditability and bounded system responses.
Learning Task Mixtures for Instruction Tuning: An explicit bilevel meta-learning strategy named ADAPT continuously learns a distribution over tasks to optimally allocate limited instruction-tuning tokens subject to explicit budgets (Kadasi et al., 4 Dec 2025). Using a smooth worst-case validation objective coupled with entropy regularization, this approach focuses training on challenging, generalization-critical tasks, consistently outperforming or matching strong static mixtures with 3–20× fewer tokens.
Adaptive Preference-Eliciting LLM Agents: ADAPT defines a benchmark and learning approach (Reflection-DPO) for sequentially uncovering and adhering to user preferences in under-specified long-horizon tasks, via active questioning (Patel et al., 5 Apr 2025). Reflection-DPO finetunes LLMs to dynamically choose between asking questions and acting, using a preference-optimal loss based on preference satisfaction rates, substantially outperforming baselines on unseen users.

7. Deterministic Prompt Scheduling and Semantic Alignment in Generation

Recently, ADAPT describes architectures for precise text-to-image compositional generation of rare or out-of-distribution concepts (Lee et al., 19 Mar 2026):

Attention Driven Adaptive Prompt Scheduling: Schedules the introduction of rare or compositional prompt tokens deterministically based on real-time attention score convergence, rather than LLM-driven heuristic switching.
Orthogonal Complement Interpolations and Latent Space Manipulation: Injects disentangled “rare concept” directions derived from CLIP embeddings directly into the semantic conditioning pipeline, enabling precise, consistent synthesis of rare attribute combinations in Stable Diffusion 3, as demonstrated empirically on RareBench.

In summary, ADAPT designates a diverse class of adaptive or adaptivity-driven systems across deep learning, robotics, multimodal processing, human-agent interaction, autoscaling, and scientific management. These systems collectively advance the state of the art by making adaptivity—across perception, action, representation, or governance—a first-class control variable, with consistent evidence of superior robustness, efficiency, and practical utility relative to fixed or naive baselines (Shao et al., 17 Mar 2026, Lyu et al., 15 Jun 2026, Wei et al., 2023, Chen et al., 16 Apr 2026, Baghel, 15 May 2026, Aydemir et al., 2023, Manik et al., 5 Apr 2026, Kadasi et al., 4 Dec 2025, Mordacq et al., 2024, Danopoulos et al., 2022, Patel et al., 5 Apr 2025, Lee et al., 19 Mar 2026, Jin et al., 2023, Lin et al., 2022).