Task Fetching Unit

• Task Fetching Unit is a modular system integrating perceptual, planning, execution, and evaluation components to perform complex fetching tasks under variable conditions.
• It employs MDP formulations, risk-sensitive learning, and multimodal fusion to optimize latency-buffer tradeoffs and processing efficiency in diverse applications.
• TFUs are implemented across robotics, mobile computing, serverless scheduling, and vehicular edge computing, adapting dynamically to real-world uncertainties.

A Task Fetching Unit (TFU) is a modular system designed to execute complex fetching tasks in response to scheduling logic, resource constraints, user instructions, or real-world uncertainties. TFUs have been implemented in domains ranging from robotics and edge computing to serverless scheduling and mobile systems. These units typically integrate perceptual, planning, execution, and evaluation components, often utilizing multimodal fusion and dynamic reliability-aware or optimal control algorithms.

1. Fundamental Architectures of Task Fetching Units

TFUs are instantiated across multiple application domains, each with a distinctive architectural blueprint.

• Mobile Computing (Prefetching): The TFU comprises a central server queue (Q₁), a mobile device buffer (Q₂), interfaces for wireless channel state monitoring, and a processor congestion model. The unit is responsible for making dynamic fetch decisions (FE/NE) per time slot, based on instantaneously observed channel and processor statistics. The core operation is a Markov Decision Process (MDP) with state $\mathbf{s}=(b_1,b_2,j,m)$ for optimal control of latency and buffer occupancy, governed by Bellman equations (0912.5269).
• Serverless Computing (Scheduling): In Hiku, the TFU is distributed: each worker runs a lightweight agent to issue IdleSignal and EvictSignal messages, while the central scheduler maintains per-function priority queues (PQ_f), sorting idle workers by load for optimal locality and response time. Assignment of new requests uses a pull-based, event-driven algorithm with fallback to least-connection scheduling (Akbari et al., 21 Feb 2025).
• Robot Manipulation (Perceptual-Action Loop): For robotic fetching, as evaluated in FetchBench and MTCM-AB, the TFU is a multimodal pipeline: perception modules extract vision and language features, motion planners compute approach/retrieval trajectories, and behavioral policies execute grasp and movement commands. Expert data generation, modular baselines (sense–plan–act, end-to-end), and risk/ambiguity handling are inherent (Magassouba et al., 2019, Han et al., 2024).
• Vehicular Edge Computing (Risk-sensitive Fetching/Offloading): The vehicle-side TFU interfaces with multicamera streams, local computation, and edge server offloading, dynamically choosing between local synthesis and remote fetch, using a learning-based, risk-sensitive entropic criterion and policy estimation updates (Batewela et al., 2019).
• Fetch-and-Carry Automation: In FCOG tasks, the TFU pipeline includes task generation (instruction synthesis, scene sampling), object grounding via NL/vision fusion, execution modules for identification, navigation, manipulation, and handover, plus unified, continuous-space evaluation (Kambara et al., 2023).

2. Computational Models and Algorithmic Formulations

TFUs universally deploy algorithmic frameworks to optimize performance targets.

• MDP/DP Formulation: Task fetching is formalized as a stochastic shortest-path MDP, with no discounting. The per-stage cost is $c(s,a)=b_1 + c\,b_2$, and the Bellman equation for expected total holding cost is:

$$V(b_1,b_2,j,m) = \min_{a\in\{\mathrm{FE},\mathrm{NE}\}} \left\{ (b_1 + c\,b_2) + \mathbb{E}[ V(\cdot) \mid a ] \right\}.$$

Value iteration and look-up policies are standard; switchover curves $\varphi(b_1)$ characterize the optimal prefetching boundaries (0912.5269).

• Attention-Based Multimodal Classification: Robotic TFUs utilize ABN-attention models; inputs include tokenized language, target and context image crops, and geometric attributes. Parallel attention branches mask salient linguistic and visual features; fusion is achieved via joint MLP embeddings, cosine similarity triplet losses, and cross-entropy for source region classification (Magassouba et al., 2019).
• Risk-Sensitive Learning: Vehicular TFUs seek to minimize entropic risk of end-to-end latency, $\frac{1}{\rho}\ln(\mathbb{E}[e^{\rho T}])$, explicitly accounting for mean, variance, skewness. A joint utility/regret policy estimation algorithm iteratively updates action probabilities in response to observed delays (Batewela et al., 2019).
• Pull-Based Scheduling: Function assignment in serverless TFUs executes:
  1. $f \leftarrow f(r)$
  2. If $PQ_f$ not empty: assign to least-loaded idle worker; else select among minima of $Load(w)$ randomly. After execution, IdleSignal and EvictSignal update PQ_f state. Mathematical models for latency, cold start rate, and load coefficient of variation (CV) define measurable impacts (Akbari et al., 21 Feb 2025).

• Scene-Parameterized Sampling & Evaluation: FetchBench task definition is parameterized by states $s\in S$, actions $a\in A$, and reward $R(s,a)$. Procedural scene generators and expert planners (CuRobo, ContactGraspNet) support automated dataset construction and baseline evaluation (Han et al., 2024).

3. Integration of Perception, Semantic Understanding, and Planning

State-of-the-art TFUs achieve fetch tasks by tightly coupling perception and semantic modules with robust planning and execution engines:

• Vision-Language Fusion: Robotic TFUs ingest user instructions (spoken/written), process via pretrained BERT and CNN encoders, and perform object grounding and semantic command parsing. Contextual attention supports ambiguity resolution and clarification through dialogue (Magassouba et al., 2019).
• Segmentation and Motion Planning: Perception employs RGB-D segmentation (e.g., SAM), generating point clouds for background and targets. Grasp pose prediction leverages ContactGraspNet, followed by trajectory planning via CuRobo or RRT-Connect. Feedback control and abort conditions handle execution errors (Han et al., 2024).
• Procedural Metric Sampling: FetchBench benchmarks diversify scenes using randomization over hundreds of parameters; object placements and support surfaces ensure a wide coverage of manipulation challenges (Han et al., 2024).
• Automated Task Generation: Fully automated TFUs synthesize natural language tasks conditioned on spatial attributes using models such as Case-Relation Transformer, promoting diversity through gaussian jittering of coordinates (Kambara et al., 2023).

4. Performance Metrics, Evaluation, and Empirical Results

TFUs are assessed via quantitative, statistically rigorous metrics.

• Mobile Systems: Key metrics include average task latency $\overline{d}$, buffer occupancy $\overline{b}_2$, and energy cost $\overline{e}$. Optimal, FON, and RFON policies achieve tradeoff curves with FON within $1$–$5$\% and RFON within $2$–$8$\% of OPT in latency-buffer tradeoff. "Always-fetch" and "never-fetch" yield opposing extremes (0912.5269).
• Serverless Platforms: Hiku's TFU realized:
  • 14.9% lower average latency (pull-based: 481 ms vs. 565–660 ms for baselines)
  • 36.4% reduction in 99th-percentile tail latency
  • cold start rate at 30% (vs. 43–59%)
  • 8.3–32.8% throughput improvement
  • load CV reduced from 0.31 to 0.27
  • strong concurrency scaling at 100 VUs (78 rps vs. 51.2–69 rps) (Akbari et al., 21 Feb 2025).
• Robotic Fetching: FetchBench baselines demonstrate low absolute success rates (maximum 20.2%). Table-top scenes are more tractable (~30–40%), while shelf/drawer/basket scenarios drop below 15%. Oracle components dramatically boost success, implicating scene completion and grasp prediction as major bottlenecks (Han et al., 2024).
• Multimodal Classification: MTCM-AB achieved 90.1 ± 0.5% top-1 accuracy on PFN-PIC (human: 90.3%), 89.2 ± 1.3% on WRS-PV (human: 94.3%), outperforming prior methods (Magassouba et al., 2019).
• Vehicular Edge: Risk-sensitive TFUs decreased delay variance by up to 40% relative to average-based baselines, produced steeper tail decay in completion time distribution, and maintained convergence with mild mean delay penalties (Batewela et al., 2019).
• Fully Automated Task Management: FCOG TFUs (Kambara et al., 2023) yielded a 20% object grounding accuracy (OLR) and 100% fetch reliability given correct grounding, but only 12.5% success in carry (handover), demonstrating high dependency on cumulative spatial errors.

5. Design Insights, Bottlenecks, and Recommendations

Cross-domain analysis identifies system-level tradeoffs, failure modes, and improvement strategies:

• Latency–Buffer Tradeoff: The cost coefficient $c$ directly modulates prefetch aggressiveness vs. memory pressure. Switchover boundaries for fetching shrink as $c$ increases, and grow with both backlog and processor service rate (0912.5269).
• Ambiguity and Clarification: Multimodal attention and semantic parsing, combined with confidence thresholds, facilitate robust handling of ambiguous instructions and user-driven error correction (Magassouba et al., 2019).
• Perceptual Limitations: Partial scene observability generates substantial failures in planning and manipulation; improving fusion of multi-view point clouds or implicit geometry modeling is critical (Han et al., 2024).
• Hybrid Strategies: Combining grasp-pose predictors for initial approach with imitation-learned retrieval policies circumvents planning failures in cluttered scenes. Re-try strategies offer only moderate benefit at increased cost (Han et al., 2024).
• Learning and Policy Adaptation: Vehicular TFUs use regret-based learning with tunable risk parameters ($\rho$, $\xi$) for mean vs. tail reliability, maintaining computational efficiency ($O(1)$ per slot) (Batewela et al., 2019).
• Scalability and Fault Tolerance: Serverless TFUs in Hiku avoid centralized heartbeat floods, allowing event-driven PQ updates and sharded scheduling for robust scalability and recovery (Akbari et al., 21 Feb 2025).

6. Domain-Specific Implementations and Future Directions

TFUs continue to evolve in response to the increasing demands of real-world systems:

• In legged robot manipulation (Helpful DoggyBot), TFUs integrate hardware grippers, agile whole-body depth-trained controllers, and pre-trained vision-LLMs for semantic generalization: demonstrated zero-shot fetching of objects in novel indoor environments with a 60% success rate, including traversal scenarios such as climbing over a queen-sized bed (Wu et al., 2024).
• Scaling up expert datasets via procedural scenes and aggressive data augmentation (domain randomization) can enhance generalization in robotic TFUs (Han et al., 2024).
• Diffusion-based policies and neural collision cost modeling in task and motion planning (TAMP) pipelines are suggested to push flexible manipulation and collision-aware planning to new heights (Han et al., 2024).
• TFUs will increasingly rely on tightly integrated multimodal fusion, risk-sensitive and adaptive scheduling, and hybrid control policies to meet the latency, reliability, and autonomy requirements of next-generation systems.