Chapter 4: How Do We Fit a New Robot and Task? — Fine-Tuning, Action Heads, and Embodiment Adaptation

Overview

Adaptation is not merely adding new data to a pretrained checkpoint. A new camera changes optics and timestamps; a new arm changes frames, workspace, and control modes; and a new gripper may assign a different physical meaning to the same scalar. The interfaces around the model therefore change with its weights. The deployable unit is a compatible tuple of base model, preprocessing, input mapping, action head, action mapping, calibration, controller, evaluator, and safety configuration.

S11 contributes the verified cell identity: robot, cameras, clocks, frames, calibration, controller, collision and force limits, watchdog, protective stop, E-stop, and accountable human. S12 contributes the versioned task and skill APIs, action semantics, proposal → projected → sent → accepted/executed lineage, feasibility projection, classical and learned baselines, independent evaluator, and rollback contract. This chapter asks a new S13 question: while holding that lower contract fixed, what should be frozen or learned for a new task or robot, and how can an interface defect be distinguished from a weight defect?

After reading this chapter... - choose among full/partial tuning, parameter-efficient updates, prompts, new heads, distillation, and retrieval on matched axes; - audit camera, frame, control-mode, action-rate, and gripper-semantics bridges independently; - separate task transfer from cross-embodiment transfer and leave unsupported hardware cells empty; - produce a tabletop new-robot adaptation runbook and an exact reversible receipt; and - diagnose regression, negative transfer, contact failure, and latency in causal order.

The experiment question is: with assembly success, initial-state distribution, evaluator, lower controller, safety limits, and target-robot data budget fixed, does adapting a pretrained policy beat a target-only S12 policy on new instructions, layouts, objects, or a bounded embodiment change—and can that gain be reproduced and rolled back exactly?

1. Mental model: adaptation is a versioned interface change

Full fine-tuning updates most or all of the backbone and head. Partial fine-tuning opens selected late blocks, normalization layers, or projections. Parameter-efficient fine-tuning freezes the base and trains a smaller delta. An adapter inserts bottleneck modules between layers; LoRA represents weight updates with low-rank matrices ^{[5] [7]}. Those foundational studies are from language tasks, so they motivate implementation candidates rather than directly establishing robot-control stability.

Prompt conditioning selects behavior through language, goal images, robot identifiers, or task tokens. A new action head maps shared features into the target robot’s action dimension, distribution, and horizon. Policy distillation transfers outputs from a larger teacher or ensemble into a student. Retrieval augmentation supplies nearby verified episodes, skill examples, or failure records as additional context. The packet lacks a fixed-budget real-assembly comparison for the last two methods, so they remain separate experiment cells rather than default promotion paths.

new task or new robot
        ↓
difference ledger: observation | frame | action | time | contact | task
        ↓
compatibility bridges ── close the gate if unresolved
        ↓
smallest learned change: condition → head → adapter → partial → full
        ↓
replay → simulation → shadow → human-reviewed bounded trial
        ↓
independent evaluation + regression decision + full-tuple rollback

Outside every stage: collision/force gates · watchdog · protective stop · E-stop · human authority

Full tuning, partial tuning, adapters, prompts, heads, distillation, and retrieval require separate regression and rollback evidence. The supporting method lineages change different state and have different failure radii ^{[1] [5] [7] [12]}. Adapter and LoRA lineage is primarily linguistic, while MAML’s robot evidence does not match this assembly cell; none supports “few trainable parameters means safe, successful transfer.” A method name can describe storage cost without proving camera, action, or contact compatibility.

Starting with the smallest change does not mean the smallest method always wins. If only output dimensions differ, a new head is the smallest coherent change. If a wrist camera is new and the backbone cannot consume it, the input projection or relevant blocks must open. If control changes from pose deltas to a force/impedance skill call, a low-rank update cannot invent the missing semantics. Specify the output contract and validate the safe executor first.

2. Make task and embodiment transfer different experiments

Task transfer keeps the robot, cameras, frames, actions, and controller fixed while changing parts, layouts, instructions, or assembly sequence. Cross-embodiment transfer changes one or more of kinematics, camera placement, workspace, action dimension, control mode, gripper, and contact geometry. Combining both under one “novel” denominator makes attribution impossible.

Task transfer and embodiment transfer are different tests; the latter audits camera, frame, action rate, controller, gripper, and contact geometry. Open X-Embodiment’s normalization/de-normalization and Octo’s flexible I/O support this audit structure ^{[11] [12]}, but they do not provide a universal adapter. Dataset-level normalization in particular does not preserve every force, compliance, timing, queue, and contact meaning.

RoboNet shows why multi-robot data can support shared learning, but it still needs platform-specific camera and action handling and mainly studies short-horizon pushing ^[6]. Bridge Data provides scene and task diversity mostly within a common embodiment, so morphology transfer remains limited ^[8]. One-Shot Visual Imitation Learning adapts within shared task structure and embodiment; one demonstration does not resolve unseen action semantics ^[2]. “Few demonstrations” is not synonymous with “few interface differences.”

3. Validate five bridges before changing weights

3.1 Camera bridge

Pin each scene/wrist camera’s name, optical frame, resolution, crop, distortion correction, channel order, exposure, and timestamp. If the pretrained policy lacked a wrist image, do not silently duplicate another view; record a changed input design. Visual domain randomization can probe background and lighting variation, but it does not transfer dynamics or contact and its bounds must cover the actual optics ^[3].

3.2 Frame bridge

Place an explicit transform between model output and the skill API. Record translation units, rotation representation, delta-composition order, reference frame, tool center point, and application time. Known-pose round trips, axis-unit motions, and zero-motion replay must pass before training. A policy can memorize a calibration bias on one layout and then fail immediately after recalibration.

3.3 Control-mode bridge

Joint position, joint velocity, Cartesian delta, and skill call are different contracts. If a model emits pose deltas but the cell requires an impedance skill, a bridge constructs bounded references and deadlines; the lower controller retains force, compliance, and stop authority. Do not add a path from model output directly to torque or force limits.

3.4 Action-rate bridge

Separate dataset sampling rate, model call rate, prediction horizon, executed prefix, and controller rate. Repetition or interpolation is valid only after speed, acceleration, freshness, and contact-transition tests. A longer prefix may be useful in free space; after the pre-contact stop, shorten execution and react to new force events.

3.5 Gripper-semantics bridge

Determine whether 0/1 means open/closed, width in meters, force request, or velocity. Do not place a parallel-gripper width and a dexterous-hand joint vector under one normalized “gripper” field. Completion uses independent width/current/force/slip and task-state observations, not the command alone. Unknown signs or ranges close the gate.

4. Choose what to freeze and train through gates

Do not decide from demonstration count alone. First determine whether observation and action meanings are compatible. Then probe whether frozen features separate target states using a linear probe, nearest-neighbor retrieval, and goal/distractor minimal pairs. If representation is adequate and only output dimension differs, train the head. If a new camera or state is essential, open its projection and relevant blocks. Consider full tuning only with enough target data and a regression suite that can measure lost prior abilities.

Within its own Franka tabletop study, OpenVLA reports that LoRA can approach full-tuning performance while training a small share of parameters ^[13]. That is a reason to include LoRA in the first matched comparison, not a rule for other robots, contact modes, and cameras. Octo was designed for flexible robot inputs and outputs but still notes limits around wrist-camera support and optimal-demonstration-heavy data ^[12].

4.1 Do not read a compound recipe as one causal effect

OpenVLA-OFT reports LIBERO success rising from 76.5% to 97.1% and 26x action-generation throughput while changing several recipe components together. The cited experiment is a simulation result, not an isolated causal factor ^[15]. It cannot be assigned solely to a head, continuous actions, parallel decoding, or tuning method, nor transferred as expected real-assembly success or end-to-end action age. A local study needs component ablations and timing from camera event through controller application.

Use these cumulative gates:

1. G0 — Contract completeness: every camera, frame, action, rate, and gripper field is compatible or resolved by an explicit bridge.
2. G1 — No-adaptation baselines: run the pretrained policy, S12 target-only policy, and classical skills on the same replay set.
3. G2 — Representation adequacy: frozen features separate target object/state/stage and reject missing or occluded evidence.
4. G3 — Smallest matched change: compare condition, head, adapter, partial, and full under the same target data, compute, and seeds.
5. G4 — Regression: replay original tasks, new tasks, failures, interventions, and safety-rejection tests.
6. G5 — Evidence promotion: keep replay, controller stub, simulation, and shadow results distinct.
7. G6 — Bounded-trial card: independent evaluation and safety owners approve the full tuple and immediate rollback.

When information is absent, record missing; when no measurement exists, not measured; when a fault test was skipped, not exercised; and when meanings disagree, incompatible. Do not convert any of these into “training will fix it.”

5. New-robot adaptation runbook for tabletop assembly

The through-task remains recognize part/fixture → grasp → collision-free transport → pre-contact stop → bounded placement/insertion → detect failure → retreat/retry/request human/stop. The first embodiment experiment compares the default single arm and parallel gripper with another single arm and parallel gripper under the same success definition and initial states. Bimanual, dexterous, mobile, and humanoid systems are advanced branches with new denominators and safety validation.

Step A — Seal invariants

Freeze task version, assets, initial-state strata, success/failure evaluator, retry budget, and human-intervention policy. Hash the S11 cell calibration/controller/safety identities and S12 skill API, evaluator, and baselines. If the new robot’s verified workspace or force envelope cannot perform the task, record a scope change and stop; learning is not the remedy.

Step B — Complete the embodiment difference ledger

Compare source and target robots on cameras, state, frames, DoF, action, rate, control mode, gripper, contact geometry, transport, and queue semantics. Label each row same, mapped, missing, or incompatible, with owner and validation test. Version transformation code separately from model code.

Step C — Seal the data budget and split

Record exposure by success, failure, intervention, recovery, object, layout, occlusion, and contact state—not only demonstration count. Keep source-robot, target-train, validation, and final-test episodes lineage-separated; compute normalization on training only. Fix physical exposure and compute so a method with more target data does not masquerade as a better adaptation algorithm.

Step D — Run the smallest-change ladder

Evaluate no adaptation, conditioning, new head, adapter/LoRA, partial tuning, and full tuning in order. Replay original and new tasks with identical evaluation IDs, seeds, and initialization. Once a stage meets the target, proceed to a larger change only if it improves cost or regression. Keep distillation and retrieval on separate branches with teacher/index versions and added latency included.

Step E — Pass the execution stack

Give raw candidates, mapped candidates, feasibility projections, sent commands, and accepted/executed references distinct IDs. Measure rejection and projection delta by free-space, pre-contact, contact, and recovery phase. If projection makes large corrections before success, credit the hybrid stack rather than the model alone. Contact retains the approved impedance/force controller and independent limits.

Step F — Inject regressions and faults

Inject camera delay, occlusion, reordering, wrong calibration generation, missing frame, action-rate change, reversed gripper sign, communication loss, expired chunks, unfamiliar handles, and contact jams. Independent gates must reject unsafe candidates even when model confidence is high. After the retry budget, transition to retreat, handoff, or stop instead of repeating the same action.

Step G — Promote or roll back exactly

Passing shadow mode confers no command authority. Independent owners review the bounded-trial workspace, speed/force envelope, episode count, stop conditions, and rollback tuple. On a gate failure or distribution change, atomically restore the last approved compatible tuple, not just an older weight file.

6. Read published cases as staged evidence and counterevidence

π0.5 open-world claims are bounded by its mobile-manipulation evaluation and reported unfamiliar-handle, occlusion, and distraction failures. The primary source documents those limits alongside its evaluation ^[14]. This supports diverse co-training as a promising direction while directly countering “open-world adaptation is solved.” The S13 single-arm insertion study must not import its success denominator; it can reuse handles, occlusion, and repeated-subtask behavior as fault-injection categories.

GR00T N1 is a humanoid advanced branch and its source focuses mainly on short-horizon tabletop manipulation. Its simulation and real-robot evaluations support that bounded reading ^[16]. They are useful for humanoid representation and synthetic-data research, but do not establish the single-arm default or validate long-horizon loco-manipulation or assembly. The real-world denominator remains source-specific GR-1 tabletop manipulation with a whole-body IK controller; adding balance, locomotion, or longer bimanual coordination requires new controllers, monitors, and failure denominators.

MAML targets rapid task adaptation from few examples, but second-order cost and negative transfer outside shared structure remain ^[1]. Dynamics randomization has useful real-robot transfer evidence, yet out-of-range contact parameters and conservative behavior remain failure modes ^[4]. RMA is real-hardware evidence for latent online adaptation, but quadruped locomotion with privileged simulation context is not tabletop insertion evidence ^[9].

Robomimic provides a manipulation counterexample: data quality, observations, algorithm, and training choices interact under controlled evaluation ^[10]. A large pretrained backbone does not remove offline-imitation bias, reset cost, or intervention cost. A close retrieval match or smooth teacher output likewise does not replace an independent evaluator.

7. Diagnose failures before blaming weights

Use this causal order: observation → frame → action meaning → time → lower control → evaluation → model. Check units and calibration generation before retraining a head. Low training loss with high projection rejection usually points to an output-contract problem. Good offline metrics with late shadow actions require a smaller model, different batching, or a local path—not model ownership of the controller rate.

Negative transfer is visible only when new-task gains and old-task losses are reported together. Do not promote an improved mean if rare occlusion, recovery, or prohibited-request rejection regresses. When a new-robot success depends on large classical projection, report the raw/projected gap instead of attributing the outcome solely to transfer.

8. Reversible adaptation receipt

A rollback receipt binds base model, adapter/head, preprocessing, calibration, action mapping, controller, evaluator, and safety versions. Model and deployment sources motivate component identity and independent monitoring ^{[12] [13] [17]}, but no model paper supplies the complete tuple. The format below is a system artifact added by the S13 execution and safety contract. If one field is missing, “roll back the model” is not yet an executable plan.

adaptation_receipt:
  receipt_id: s13-ch04-<run>-<sha>
  base_model: {checkpoint: <sha>, code: <commit>, license: <id>}
  learned_delta: {method: <head|adapter|partial|full|distill|retrieval>, artifact: <sha>}
  freeze_map: {trainable_modules: [...], frozen_modules: [...], optimizer_state: <sha>}
  data: {manifest: <sha>, train_val_test: <ids>, normalization_split: <id>, budget: <value>}
  observations: {camera_contract: <sha>, preprocessing: <sha>, state_schema: <sha>}
  geometry: {tf_tree: <sha>, calibration_generation: <id>, tool_frame: <id>}
  actions: {schema: <sha>, mapping: <sha>, rate_hz: <value>, gripper_semantics: <id>}
  execution: {task: <ver>, skill_api: <ver>, projection: <ver>, controller: <ver>}
  evaluation: {protocol: <sha>, denominators: <artifact>, independent_signoff: <ids>}
  safety: {limits: <ver>, monitor: <ver>, watchdog: <ver>, stop_tests: <artifact>}
  evidence: {replay: <id>, simulation: <id>, shadow: <id>, bounded_hardware: <id|not_exercised>}
  rollback_target: {complete_compatible_tuple: <receipt_id>, restore_test: <artifact>}
  invalidation: [camera_change, calibration_change, controller_change, data_overlap_found, drift_limit]

The receipt includes failures, rejection reasons, projection deltas, human interventions, recoveries, and protective-stop events, all linked by execution identity. A restore test does more than load files: it confirms that the prior tuple recovers its I/O contract on identical replay without leaking the newer bridge. Model rollback never erases a safety event or clears a stop.

9. Metrics, artifacts, and promotion decision

The mandatory bundle is embodiment_diff_ledger, camera_contract, frame_roundtrip, action_mapping, gripper_semantics, freeze_map, training_receipt, matched_regression, execution_trace, fault_injection, adaptation_receipt, and rollback_restore_test. Keep task, data, model, robot/control, evaluation, and safety signatures independent.

9.1 Bounded Codex implementation prompt

Goal
- Compare new-robot adaptation candidates on recorded tabletop-assembly episodes.
- Evaluate no-adaptation, new-head, adapter/LoRA, partial-tune, and full-tune
  under matched data and compute budgets.
- Produce a reversible adaptation receipt and exact rollback target.

Context
- Read S11 cell/calibration/safety identities and S12 task/skill/action/controller/evaluator contracts.
- Compare source and target camera/frame/control-mode/action-rate/gripper semantics.
- Use recorded replay, controller stubs, kinematic projection, simulation, and shadow only.

Constraints
- Do not call robot drivers, send_action, controller switching, fault clearing, or limit-changing APIs.
- Preserve missing/not measured/not exercised/incompatible instead of guessing.
- Keep task transfer and embodiment transfer in different denominators.
- Do not merge raw/proposed/projected/sent/accepted-executed identifiers.
- Keep collision/force/watchdog/protective-stop/E-stop/human authority outside the model.

Done when
- embodiment_diff_ledger and every mapping round-trip test exist.
- Freeze map, dataset hashes/splits, seeds, and checkpoint/delta hashes are recorded.
- Regression is reported on old tasks, new tasks, failure/intervention/recovery slices.
- Inject stale camera, wrong calibration generation, reversed gripper, rate mismatch,
  infeasible pose, queue reordering, outage, and contact jam; verify fallback.
- Report success, rejection/projection, action age, force event, intervention,
  recovery/handoff, compute/cost, and an exact rollback restore test.

Safety
- Do not execute a hardware trial or clear a stop.
- You may only produce a separate human-reviewed bounded-trial card.
- Learned uncertainty and retrieval similarity are not safety authorization.

10. Limits, companion reading, and the Chapter 5 interface

The evidence packet does not provide a matched real-assembly comparison across every adaptation method. Adapter and LoRA foundations come from language models; MAML does not use the target assembly hardware; and the central OpenVLA-OFT result is a compound simulation recipe. Open X-Embodiment has substantial real-robot evidence, but its robot-specific evaluations, actions, and controllers remain heterogeneous. \pi_{0.5} and GR00T N1 retain their mobile-manipulation and humanoid denominators. Distillation and retrieval especially lack a fixed-budget contact-rich rollback comparison.

This chapter therefore names no universal winner. With compatible interfaces, compare the smallest learned change first; with incompatible interfaces, validate bridges and the lower executor first. Separate new-task gain from old-task regression, target-robot success from projection correction, and generation throughput from end-to-end action age. Do not promote a more complex adaptation that fails to beat the classical baseline or the target-only S12 policy.

Terry’s \pi_0 explainer, \pi_{0.5}-lineage RECAP, RoboPaint, Gen-1, Act-1, Habilis-β, UMI, UMI-FT, Octo, and RoboVLMs are companion readings, not substitutes for the primary sources below.

Chapter 5 receives more than an adapted model: versioned camera, frame, action, and gripper contracts; a freeze map; an embodiment difference ledger; a validated adaptation tuple; failure taxonomy; independent evaluation; and an immediately restorable prior tuple. Chapter 5 will ground language into objects, relations, goal states, and skill preconditions on this fixed interface. Correctly interpreting “the right hole” cannot repair a wrong frame, unit, or gripper sign. Conversely, a stable adaptation contract lets Chapter 5 reject language ambiguity and spatial-grounding defects without confounding them with embodiment errors.

Supplementary evidence map

The following verified primary-source groups come from the independent S13 ledger. They document search coverage without changing existing claims or denominators.

• composable skills: BiNoMaP: Learning Category-Level Bimanual Non-Prehensile Manipulation Primitives, RH20T-P: A Primitive-Level Robotic Dataset Towards Composable Generalization Agents, RoVi-Aug: Robot and Viewpoint Augmentation for Cross-Embodiment Robot Learning, GHIL-Glue: Hierarchical Control with Filtered Subgoal Images — source-specific tasks, data, and evidence stages remain separate.
• data mixtures and augmentation: Re-Mix: Optimizing Data Mixtures for Large Scale Imitation Learning, Whole-Body Teleoperation for Mobile Manipulation at Zero Added Cost, Diffusion Transformer Policy — source-specific tasks, data, and evidence stages remain separate.
• post-training and assembly: FLaRe: Achieving Masterful and Adaptive Robot Policies with Large-Scale Reinforcement Learning Fine-Tuning, Video Prediction Policy: A Generalist Robot Policy with Predictive Visual Representations, Towards Natural Language-Driven Assembly Using Foundation Models — source-specific tasks, data, and evidence stages remain separate.

References

1. Finn, C., Abbeel, P., & Levine, S. (2017a). Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks. ICML. arXiv:1703.03400.
2. Finn, C., Yu, T., Zhang, T., Abbeel, P., & Levine, S. (2017b). One-Shot Visual Imitation Learning via Meta-Learning. CoRL. arXiv:1703.07326.
3. Tobin, J., et al. (2017). Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World. IEEE IROS. arXiv:1703.06907.
4. Peng, X. B., Andrychowicz, M., Zaremba, W., & Abbeel, P. (2018). Sim-to-Real Transfer of Robotic Control with Dynamics Randomization. IEEE ICRA. arXiv:1710.06537.
5. Houlsby, N., et al. (2019). Parameter-Efficient Transfer Learning for NLP. ICML. arXiv:1902.00751.
6. Dasari, S., et al. (2020). RoboNet: Large-Scale Multi-Robot Learning. CoRL. arXiv:1910.11215.
7. Hu, E. J., et al. (2021). LoRA: Low-Rank Adaptation of Large Language Models. ICLR. arXiv:2106.09685.
8. Ebert, F., Nair, S., Lee, A., Levine, S., & Finn, C. (2021). Bridge Data: Boosting Generalization of Robotic Skills with Cross-Domain Datasets. CoRL. arXiv:2109.13396.
9. Kumar, A., Fu, Z., Pathak, D., & Malik, J. (2021). RMA: Rapid Motor Adaptation for Legged Robots. Robotics: Science and Systems. arXiv:2107.04034.
10. Mandlekar, A., et al. (2021). What Matters in Learning from Offline Human Demonstrations for Robot Manipulation. CoRL. arXiv:2108.03298.
11. Open X-Embodiment Collaboration. (2024). Open X-Embodiment: Robotic Learning Datasets and RT-X Models. IEEE ICRA. arXiv:2310.08864.
12. Octo Model Team, et al. (2024). Octo: An Open-Source Generalist Robot Policy. arXiv:2405.12213.
13. Kim, M. J., et al. (2024). OpenVLA: An Open-Source Vision-Language-Action Model. arXiv:2406.09246.
14. Physical Intelligence, et al. (2025). $\pi_{0.5}$: a Vision-Language-Action Model with Open-World Generalization. arXiv:2504.16054.
15. Kim, M. J., Finn, C., & Liang, P. (2025). Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. arXiv:2502.19645.
16. NVIDIA, et al. (2025). GR00T N1: An Open Foundation Model for Generalist Humanoid Robots. arXiv:2503.14734.
17. Agia, C. (2026). Deployment-Time Reliability of Learned Robot Policies. arXiv:2603.11400.
18. Zhou, H., et al. (2025). BiNoMaP: Learning Category-Level Bimanual Non-Prehensile Manipulation Primitives. arXiv primary preprint (2025-09-25). arXiv:2509.21256.
19. Chen, Z., et al. (2024). RH20T-P: A Primitive-Level Robotic Dataset Towards Composable Generalization Agents. arXiv primary preprint (2024-03-28). arXiv:2403.19622.
20. Chen, L. Y., et al. (2024). RoVi-Aug: Robot and Viewpoint Augmentation for Cross-Embodiment Robot Learning. arXiv primary preprint (2024-09-05). arXiv:2409.03403.
21. Hatch, K. B., et al. (2024). GHIL-Glue: Hierarchical Control with Filtered Subgoal Images. arXiv primary preprint (2024-10-26). arXiv:2410.20018.
22. Hejna, J., et al. (2024). Re-Mix: Optimizing Data Mixtures for Large Scale Imitation Learning. arXiv primary preprint (2024-08-26). arXiv:2408.14037.
23. Honerkamp, D., et al. (2024). Whole-Body Teleoperation for Mobile Manipulation at Zero Added Cost. IEEE Robotics and Automation Letters, vol. 10, no. 4, pp. 3198-3205, April 2025. arXiv:2409.15095.
24. Hou, Z., et al. (2024). Diffusion Transformer Policy. arXiv primary preprint (2024-10-21). arXiv:2410.15959.
25. Hu, J., et al. (2024). FLaRe: Achieving Masterful and Adaptive Robot Policies with Large-Scale Reinforcement Learning Fine-Tuning. arXiv primary preprint (2024-09-25). arXiv:2409.16578.
26. Hu, Y., et al. (2024). Video Prediction Policy: A Generalist Robot Policy with Predictive Visual Representations. arXiv primary preprint (2024-12-19). arXiv:2412.14803.
27. Joglekar, O., et al. (2024). Towards Natural Language-Driven Assembly Using Foundation Models. arXiv primary preprint (2024-06-23). arXiv:2406.16093.