Chapter 9: Build the Learning Stack One Gate at a Time — A Baseline-to-Policy Runbook

Overview

This chapter introduces no new algorithm. It turns the execution hierarchy, planning, contact control, dataset, BC/ACT, generative policy, selective RL, and async fallback contracts from Chapters 1–8 into one tabletop-assembly runbook. Comparisons retain one task contract. A stage is added only when a measured gap, a testable hypothesis, and a budget justify it.

The running task uses a 6/7-DoF arm, parallel gripper, and wrist/scene cameras to grasp an object, transport it collision-free, insert it, and retreat. A classical baseline first verifies feasibility and controller ownership. Dataset QA verifies demonstration meaning. Plain BC and ACT establish learned baselines. Diffusion or flow is considered only after measured multimodality; RL opens only after pre-registering reward-based incremental value.

Evidence boundary: The full ladder is original S12 synthesis. MoveIt, ACT, Diffusion Policy, and HIL-SERL support components, not the complete classical-to-imitation-to-generative-to-RL-to-hardware sequence ^{[12] [7] [8] [11]}. The Capuano tutorial supplies implementation orientation and an API-reading aid, never a code dump or hardware authorization path ^[14].

After reading this chapter... - You can freeze a versioned experiment card and task/action/controller/evaluator contract. - You can distinguish done-when evidence for classical, data, BC/ACT, generative, and optional-RL gates. - You can compare denominator, intervention, safety events, latency, recovery, drift, and shift. - You can separate simulation, shadow, and bounded-hardware evidence. - You can prepare rollback and release evidence. - You can hand #S13 stable skill, data, evaluation, and safety contracts rather than only weights.

The final question is: What is the simplest stack that satisfies pre-registered KPIs on the same assembly task, and does each added policy justify its data, controller, compute, human, and hardware cost?

1. Gate 0 — Freeze the experiment card

Experiment identity hashes task@version + dataset + checkpoint + code + config + controller + evaluator. A screenshot or weight filename is not identity. A configuration change creates a new row and never overwrites an earlier result.

1.1 Common metric contract

Every stage preserves artifact identity, success denominator, intervention, collision/force events, latency, recovery, calibration drift, and distribution shift. The full set is S12 synthesis; official LeRobot documentation verifies only its documented episode/evaluation workflow ^{[15] [16] [10]}.

$$

\hat p=N_{success}/N_{eligible},\quad

r_{int}=N_{intervention}/N_{eligible},\quad

r_{safe}=(N_{collision}+N_{force}+N_{stop})/N_{eligible}.

$$

Freeze eligible before results. Classify setup faults, missing sensors, operator aborts, protective stops, and policy timeouts under pre-registered inclusion rules. Separate unassisted from intervention-assisted success.

2. Gate 1 — Deterministic classical baseline

Connect grasp candidates, IK, collision-aware paths, time parameterization, trajectory execution, and contact control. RRT-Connect and TOPP-RA provide path-planning and path-timing lineages ^{[1] [5]}. They are not universal winners; version planner seeds, timeout, scene, and limits.

MoveIt connects planning-scene state to controller actions and separates geometric planning from timing ^{[12] [13]}. ros2_control and the hardware controller own execution boundaries ^{[19] [20]}.

Done when

• The phase FSM and owner–rate–deadline–failure matrix exist.
• Start/goal validity, collisions, IK branch, timing, and tolerances are logged.
• Free-space and contact evaluators are separate.
• Denominator and failure taxonomy are stable across initial bins.
• Hold, retreat, watchdog, human stop, and rollback pass fault injection.
• Planner, trajectory, and controller artifacts replay deterministically.

Do not conceal calibration, geometry, timing, or contact defects with learning. Learned baselines inherit the same action and controller boundary.

3. Gate 2 — Dataset release and QA

Preserve raw/mapped/sent actions, event/arrival times, calibration epoch, interventions, reset, failure, and safety events. UMI and DROID represent distinct collection lineages ^{[9] [10]}, not automatic validation of this dataset.

Done when

• Task card and episode termination agree.
• Object/scene/operator/calibration split leakage is absent.
• Sensor/action age and clock uncertainty meet thresholds.
• Processor unit, sign, frame, order, saturation, and round-trip tests pass.
• Replay detects missing, drop, reorder, and impossible velocities.
• Intervention authority and recovery targets remain distinguishable.
• Release manifest, parent, hashes, and quarantine/reject lists exist.

If QA fails, repair recording, mapping, or calibration before collecting more episodes.

4. Gate 3 — Plain BC and ACT

Plain BC tests semantics and closed-loop shift ^{[2] [3]}. RoboMimic is a lineage for examining data, observation, and architecture choices ^[6]. ACT adds action chunks and temporal aggregation ^[7].

Done when

• Dataset, splits, processors, configs, and checkpoints are hashed.
• Phase errors, chunk boundaries, and projection deltas are reported.
• Delay, drop, and contact simulation regressions pass.
• Shadow compares proposed, projected, and sent actions.
• Inference tails, action age, and queues satisfy deadlines.
• ACT improvement is attributed beyond plain chunking.
• Rollback to classical or the previous checkpoint is deterministic.

A time-limited LeRobot rollout is not a motion-safety envelope or sufficient evaluation protocol ^[16]. Workspace, speed, force, owners, stop paths, and denominator remain separate.

5. Gate 4 — Diffusion or flow only for measured multimodality

Skip this gate when ACT suffices. Open it when the same observation truly admits several valid trajectories and deterministic chunks show mean actions, mode omission, or recovery failure. Diffusion Policy is the primary reference for conditional denoising action sequences and receding-horizon execution ^[8]. A flow candidate must meet the same task, solver, timing, and feasibility contract.

Done when

• Multimodality is separated from hidden phase, label noise, and clock defects.
• ACT and the candidate share data, horizons, controller, and evaluator.
• Sampling/solver steps and end-to-end latency are recorded.
• Diversity and task validity are evaluated together.
• Feasibility reruns after trim or blending.
• Deadline, stale, OOD, and recovery stress tests pass.
• Benefit exceeds added compute and operational complexity.

Ten real episodes in an official contribution workflow are only a coarse starting point, not statistical sufficiency for variable tasks ^[15]. Pre-specify denominator, initial bins, failure taxonomy, and uncertainty.

6. Gate 5 — Selective RL is optional

Run RL only when a matched BC/ACT baseline leaves a measurable reward-addressable gap worth reward, reset, human, and hardware cost. This is S12 promotion policy; unmatched ACT, BAKU, and HIL-SERL numbers are not ranked ^{[7] [11]}.

The RL card records failure slice, reward/evaluator separation, reset owner, interaction cap, human minutes, exposure, branch type, and a NO_RL exit. Begin with replay/offline, then simulation, then shadow. Online real-robot RL is never a reader requirement.

Done when

• Incremental benefit and safety non-inferiority are pre-registered.
• ACT retraining and controller-tuning ablations exist.
• Reward challenge and offline-support audit pass.
• Reset distribution and manual labor are logged.
• Projection, watchdog, E-stop, and human authority remain external.
• Cost-inclusive evidence selects PROMOTE, REVISE, or NO_RL.

7. Gates 6–8 — Simulation, shadow, bounded hardware

MuJoCo can support deterministic replay and stress testing ^[4], but simulation success is not real-contact or safety evidence. Separate nominal and held-out geometry, dynamics, contact, sensor, latency, and distribution stress.

In shadow mode, the approved stack moves while the candidate proposes. Join proposal, projection, sent action, evaluator, age, queue, fallback, and authority under one cycle_id. Systematic rejection or stale/phase conflict blocks hardware.

A bounded trial uses more than a time limit. Its signed card fixes workspace, speed, force, mode, maximum episodes/interactions, stop and reset owners, abort thresholds, independent E-stop, and rollback checkpoint. Never tune configuration or reward inside a block.

8. Worked end-to-end trace

Do not beautify this episode into success or delete it. Use the trace to attribute planner, policy, controller, or network causes, then rerun the same initial bin and evaluator.

9. Failure diagnosis and cross-chapter checklist

Every row retains task/dataset/checkpoint/config/controller/evaluator hashes, denominator, success, intervention, collision/force/stop, latency, recovery, drift, and shift. Missing is not measured, never zero.

Release evidence includes reproducible build/config locks; claim/source/factcheck trail; simulation, shadow, and bounded reports; checkpoint allowlist and rollback; figure provenance; KO/EN build/link validation; and a release receipt only after passing full QA.

9.1 Configuration and evidence bundle

Close each gate with a machine-readable manifest and a human-readable verdict. The manifest records parent hashes and schema versions. The report explains pass or failure and links immutable raw evidence. Classical evidence combines planner, controller, and FSM configuration with state, path, trajectory, and fault logs. Dataset evidence combines recorder, schema, splits, immutable episodes, replay, and quarantine decisions. Learned-policy evidence combines training and processor configuration, curves, checkpoints, predictions, timing, and matched evaluation. Runtime evidence combines queue, FSM, and fallback configuration with packet, clock, transition, and fault-injection traces. Hardware evidence combines the signed card, raw cycles, incidents, and blinded evaluator output.

Do not inherit defaults silently. Distinguish absent, null, and inherited values. Reproducibility includes dependency locks, CPU/GPU, math flags, driver and firmware, camera exposure, network topology, and controller rate—not only a seed. A weight is not deployable without compatible processors, action schema, controller mode, queue policy, fallback, and allowlist.

9.2 Error budget

Separate task error into perception, calibration, planning/discretization, controller tracking, contact/model, and latency-conditioned policy error. Do not add them into one universal scalar because phases, units, and correlations differ. Give every component a probe, frame, uncertainty, owner, and threshold.

A camera change can reduce perception error while increasing decode latency. A smoother trajectory can lower tracking error and lengthen cycle time. A larger model can improve offline error while consuming deadline margin. Preserve these trade-offs instead of hiding them in one score. Retrain only after upstream errors meet their gate thresholds.

9.3 Rollback rehearsal

Rollback is not merely loading an older weight. Atomically choose a compatible checkpoint, processor, schema, controller mode, queue epoch, and FSM state; purge old chunks; and verify physical robot state. Rolling back during contact cannot assume the robot returned to the staging pose.

Rehearse inference crash, corrupt checkpoint, schema mismatch, clock jump, network disconnect, projection storm, controller rejection, and human takeover. Each scenario ends in one defined state: SAFE_HOLD, RETREAT, CLASSICAL_FALLBACK, HUMAN_RECOVERY, or STOP. An undefined transition fails the gate.

After rollback, run a baseline regression covering schema compatibility, units, controller acceptance, queue purge, evaluator behavior, and representative free-space and contact phases. Update the allowlist only after it passes. Record approver, recovery duration, and whether degraded-run data are quarantined.

9.4 Decision table

STOP_LADDER is deployment success, not research failure. Prefer the simpler row when operational evidence cannot justify added components. Do not change denominators, exclusions, tolerances, or task bins after seeing model ordering.

9.5 Release receipt and lineage audit

The robot-experiment receipt contains exact task, dataset, checkpoint, controller, and evaluator hashes; environment; denominator and outcomes; known limitations; rollback target; and owner sign-off. It links raw logs and declares exclusions. The survey receipt separately records manuscript build, references, figures, fact-checking, QA scorecard, deployment, and live bilingual verification. A blocked preview URL is not a release receipt; neither is a successful video without denominator, incidents, and rollback.

Audit parent resolution. A checkpoint points to an existing dataset and processor. A shadow report points to the exact checkpoint and controller it observed. A bounded trial points to a passing shadow report and signed card. A release receipt points to immutable reports rather than latest.

Preserve negative evidence: rejected checkpoints, quarantined episodes, missed deadlines, aborted trials, and failed rollback rehearsals. They explain why the approved row is credible and prevent survivor bias. Archive read-only commands that rebuild tables and hashes from raw logs without contacting hardware, mutating data, or fetching an unpinned model. Missing artifacts must fail loudly.

9.6 Operational handoff rehearsal

Have a second operator reproduce the report from the receipt without oral context. They should identify the approved checkpoint, start state, trial envelope, stop owner, expected telemetry, rollback command, and prohibited actions. If they must guess a device, frame, unit, controller mode, or threshold, the handoff fails.

Run the rehearsal once with the inference server unavailable and once with a deliberately incompatible schema. Both should remain in a non-moving or validated fallback state, generate bounded reason codes, and preserve evidence. This demonstrates that documentation and runtime contracts agree rather than merely existing as separate files.

Finally, expire evidence deliberately when a dependency, firmware, calibration epoch, or safety assessment changes. A receipt can remain historically valid while no longer authorizing a new trial. Record the superseding artifact and require the affected regression gates again.

9.7 Shift and maintenance triggers

Approval is not permanent after release. Maintain a trigger table for changes in object supplier, fixture, lighting, camera mount, tool, firmware, network, controller, policy dependency, or safety assessment. A calibration change may require replay and shadow. Controller firmware can reopen classical, contact, and runtime regression. An observation-schema change can invalidate every learned checkpoint back to the dataset gate.

Production monitoring tracks more than success. Compare intervention, projection, deadline miss, recovery, force events, and missing modalities by initial-state bin against the approved baseline distribution. Crossing a threshold initiates evidence freeze, safe fallback, and owner review—not automatic retraining. Quarantine failure episodes with provenance and never absorb the test population into training silently.

A maintenance receipt records the trigger, affected artifacts, temporary operating restriction, completed regressions, new approval, and superseded receipt. It also identifies whether the previous result remains scientifically reportable even though it no longer authorizes operation. This lifecycle turns the runbook from a one-off demonstration into a maintainable learning stack.

Close with a second-person audit. Another operator, without oral context, must locate the approved checkpoint, start state, trial envelope, stop owner, expected telemetry, rollback, and prohibited actions from the receipt. If they must guess a device, frame, unit, mode, or threshold, the handoff fails. Repeat with the inference server unavailable and with an incompatible schema. Both cases must remain non-moving or enter validated fallback, emit bounded reason codes, and preserve enough evidence for diagnosis. Record the rehearsal date, environment, observed transition, and remediation owner.

9.8 Freeze the release candidate and require independent sign-off

Before requesting a bounded trial, assemble one immutable release-candidate bundle. It contains more than checkpoint weights: exact hashes for processors, observation/action schemas, calibration epoch, controller mode and rate, queue and FSM configuration, evaluator, task bins, denominator, exclusion rules, rollback target, and dependency lock. The human-readable trial card and machine-readable manifest must point to the same bundle hash. Changing a threshold, tolerance, denominator, or task mix after the freeze creates a new candidate ID and reopens the affected gates; it does not silently amend the approval.

A reviewer other than the author performs a second, read-only evidence check. They reconcile raw episodes with the aggregate denominator, confirm that every exclusion retains a reason, resolve checkpoint parents to the approved dataset and processor, and verify that the rollback tuple can actually load. The review compares interventions, stops, collision or force events, deadline misses, recovery, and missing modalities by initial-state bin rather than accepting a success video or mean score. Unknown evidence remains not measured or not exercised, never an inferred pass.

Run a no-motion reconstruction dry run before sign-off. From stored packets and event logs alone, regenerate queue decisions, stale-chunk rejections, FSM transitions, evaluator verdicts, and report tables, then compare artifact hashes. Inject an unavailable inference server, incompatible schema, and missing calibration receipt. The queue must remain empty and the system must stay non-moving or enter a validated fallback. This exercise stops at the controller-interface stub and sends no hardware command.

The two signatures represent different responsibilities. The experiment owner attests that the candidate represents the intended task and operating envelope. The safety or operations reviewer approves the stop owner, prohibited actions, rollback, and evidence completeness. Neither role may waive missing evidence. Configuration drift, dependency changes, calibration expiry, or a failed rehearsal moves the candidate from FROZEN to INVALIDATED and removes trial authority. The receipt records who reviewed which evidence hashes, when the review occurred, and exactly which gates are required for reapproval.

10. Bounded Codex prompt

Goal
- Build the experiment-card validator and promotion report for the assembly
  ladder; execute no hardware motion.
Context
- Read task, dataset/split, controller, evaluator, FSM, queue/fallback,
  safety ownership, and approved checkpoint manifests.
Constraints
- Keep task/action/controller/evaluator fixed across comparison rows.
- Gate classical, data QA, BC, chunked BC, ACT, generative, optional RL,
  simulation, shadow, and bounded-trial-card review.
- Record hashes, denominator, intervention, safety events, latency, recovery,
  drift, shift, human/reset/exposure cost, and rollback for every row.
- Treat skipped gates and NO_RL as valid; invent no setting.
Done when
- Tests reject incompatible artifacts and leaked splits.
- Reports identify failed gate, owner, evidence, and rollback.
- Replay reproduces the worked trace and stale-chunk rejection.
- Output a human-reviewed trial card, not a hardware command.
Safety
- Policy/LLM/VLA never owns torque, force limits, watchdog, E-stop, or approval.

11. #S13 handoff contract

#S13 inherits versioned task/skill interfaces, dataset metadata, evaluation suites, fallback states, and independent safety contracts rather than raw weights alone. This is an editorial handoff rule, not a generalist-model performance claim ^[14].

The bundle includes schemas and lineage, controller rate/deadline, async/fallback semantics, evaluator taxonomy, approved checkpoints, promotion evidence, and unresolved gaps. #S13 re-researches VLA pretraining, action representation, cross-embodiment transfer, grounding, world models, and agents at a new cutoff. No upstream model inherits downstream safety authority.

Manufacturing Cell Checkpoint

• [ ] Gate 0 contract and budget froze before results.
• [ ] Classical baseline and rollback are validated.
• [ ] Dataset passes replay and split QA.
• [ ] Plain BC exists and ACT benefit is attributed.
• [ ] Generative/RL gates are skipped without measured need.
• [ ] Simulation, shadow, and hardware evidence remain separate.
• [ ] Independent stop authority and rollback are on the trial card.
• [ ] Denominator and taxonomy persist across every row.
• [ ] Release is declared only after full QA and receipt.

Relation to Prior Surveys

#S11 commissioning, first motion, and evidence-before-energy become Gate 0–1 preconditions rather than repeated material. Final public cross-links wait until #S11 is release-ready.

What to Learn Next

#S12 ends by selecting the simplest stack that satisfies KPIs with rollback evidence, not the most complex policy. #S13 expands upstream intelligence on top of this contract.

References

1. Kuffner, J. J., & LaValle, S. M. (2000). RRT-Connect. ICRA.
2. Argall, B. D., et al. (2009). A Framework for Learning from Demonstration. RAS.
3. Ross, S., Gordon, G., & Bagnell, J. A. (2011). A Reduction of Imitation Learning and Structured Prediction to No-Regret Online Learning. AISTATS.
4. Todorov, E., et al. (2012). MuJoCo. IROS.
5. Pham, H., & Pham, Q.-C. (2018). A New Approach to Time-Optimal Path Parameterization Based on Reachability Analysis. IEEE Transactions on Robotics. DOI: 10.1109/TRO.2018.2819195. arXiv:1707.07239.
6. Mandlekar, A., et al. (2021). RoboMimic. CoRL.
7. Zhao, T. Z., et al. (2023). Learning Fine-Grained Bimanual Manipulation. RSS.
8. Chi, C., et al. (2023). Diffusion Policy. RSS. Terry's explainer.
9. Ha, H., et al. (2024). Universal Manipulation Interface. RSS.
10. Khazatsky, A., et al. (2024). DROID. RSS.
11. Luo, J., Xu, C., Wu, J., & Levine, S. (2024). Precise and Dexterous Robotic Manipulation via Human-in-the-Loop Reinforcement Learning. arXiv:2410.21845.
12. MoveIt / PickNik (2025a). move_group Architecture. Official docs.
13. MoveIt / PickNik (2025b). Trajectory Processing. Official docs.
14. Capuano, F., et al. (2025). Robot Learning: A Tutorial. arXiv:2510.12403v1.
15. Hugging Face (2026a). Adding a Policy. Official docs.
16. Hugging Face (2026b). Policy Deployment and Rollout. Official docs.
17. Hugging Face (2026c). Human-In-the-Loop Data Collection. Official docs.
18. Hugging Face (2026d). Asynchronous Inference. Official docs.
19. ros-controls (2026a). ros2_control Architecture. Official docs.
20. ros-controls (2026b). Joint Trajectory Controller. Official docs.
21. Samad Hayati (1985). Geometric Approach for Robotic Calibration. Annotated primary reading. DOI: 10.1109/ROBOT.1985.1087372. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: Geometric calibration cannot absorb compliance, backlash, thermal drift, base motion, or time delay without corrupting parameters.
22. Christopher G. Atkeson et al. (1986). Estimation of Inertial Parameters of Manipulator Loads and Links. Annotated primary reading. DOI: 10.1177/027836498600500306. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: Numerical differentiation, sensor bias, friction, unmodeled flexibility, and closed-loop correlation can dominate estimates.
23. Maxime Gautier (1991). A Direct and Efficient Method for Identifying the Minimum Dynamic Parameters of Robots. Annotated primary reading. DOI: 10.1109/ROBOT.1991.131896. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: A minimal algebraic set may still violate physical consistency or be poorly excited by the chosen experiment.
24. Jan Swevers et al. (1997). Excitation Trajectories for the Identification of Base Inertial Parameters of Robots. Annotated primary reading. DOI: 10.1177/027836499701600306. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: Needs torque estimates and safe trajectory bounds; friction, temperature, payload, and controller dynamics require separate modeling.
25. Tobias Kunz & Mike Stilman (2012). Time-Optimal Trajectory Generation for Path Following with Bounded Acceleration and Velocity. Annotated primary reading. DOI: 10.15607/rss.2012.viii.027. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: Performance depends on geometry, initialization, objectives, and model fidelity; feasibility does not imply controller tracking.
26. Alexandre Janot et al. (2014). Identification of Robot Dynamics with the Instrumental Variable Method. Annotated primary reading. DOI: 10.1109/TRO.2014.2319567. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: Requires a useful initial model and noise assumptions; unmodeled friction and controller nonlinearities can remain.
27. Silvio Traversaro et al. (2016). Identification of Fully Physical Consistent Inertial Parameters using Optimization on Manifolds. Annotated primary reading. DOI: 10.1109/IROS.2016.7759801. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: Optimization is nonconvex and still depends on excitation, torque quality, and correct rigid-body structure.
28. Hung Pham & Quang-Cuong Pham (2018). TOPP-RA: A Fast and Robust Implementation of Time-Optimal Path Parameterization for Robots. Annotated primary reading. DOI: 10.1109/tro.2018.2821155. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: Performance depends on geometry, initialization, objectives, and model fidelity; feasibility does not imply controller tracking.
29. Patrick M. Wensing et al. (2018). Linear Matrix Inequalities for Physically Consistent Inertial Parameter Identification. Annotated primary reading. DOI: 10.1109/TRO.2017.2769099. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: Physical consistency does not guarantee identifiability, correct friction/actuator models, or transfer under different payloads.
30. Fanbo Xiang et al. (2020). SAPIEN: A SimulAted Part-based Interactive ENvironment. Annotated primary reading. DOI: 10.1109/cvpr42600.2020.01104. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: Simulation throughput or benchmark success does not establish real contact fidelity or hardware safety.
31. C. Daniel Freeman et al. (2021). Brax: A Differentiable Physics Engine for Large Scale Rigid Body Simulation. Annotated primary reading. arXiv:2106.13281. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: Simulation throughput or benchmark success does not establish real contact fidelity or hardware safety.
32. Han et al. (2024). Flange-Based 3D Hand–Eye Calibration for Soft Robotic Tactile Welding. Annotated primary reading. DOI: 10.1016/j.measurement.2024.115376. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: Welding/tactile geometry differs from tabletop RGB-D
33. Hongjie Fang et al. (2024). AirExo: Low-Cost Exoskeletons for Learning Whole-Arm Manipulation in the Wild. Annotated primary reading. arXiv:2309.14975. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: Platform, task, dataset, and evaluation assumptions must be checked in the primary paper before transferring a result.
34. Tomohiro Motoda et al. (2025). AIST Bimanual Manipulation Dataset. Annotated primary reading. canonical URL. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: ALOHA actions and camera placement are embodiment-specific
35. Yuhan Wang et al. (2025). ExDex: Dexterous Non-Prehensile Manipulation for Ungraspable Objects. Annotated primary reading. arXiv:2503.23120. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: Exact object counts and trial denominators require full-table extraction
36. Chi Zhang et al. (2025). UniTacHand: Unified Spatio-Tactile Representation for Human to Robotic Hand Skill Transfer. Annotated primary reading. arXiv:2512.21233. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: Platform, task, dataset, and evaluation assumptions must be checked in the primary paper before transferring a result.
37. Yeseung Kim et al. (2026). A Visuo-Tactile Data Collection System with Haptic Feedback for Coarse-to-Fine Imitation Learning. Annotated primary reading. arXiv:2605.08757. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: Platform, task, dataset, and evaluation assumptions must be checked in the primary paper before transferring a result.
38. Junji Oaki et al. (2026). A Reproducible and Physically Feasible Dynamic Parameter Identification Framework for a Low-Cost Robot Arm. Annotated primary reading. arXiv:2605.15949. — Role: primary reading for common task contracts, validation, and promotion artifacts. Limitation: One low-cost arm does not cover tendon hands or torque-sensor cobots