The direct answer
What is supported? The authors report a shared actuator-level policy trained in simulation on a four-cable translational point-mass CDPR. The state includes current and target cable length, cable length at t-3 and motor current. The same policy is then reported on selected eight-cable simulations and one real eight-motor translational CDPR.
What is the strongest physical result? Figure 9 is described as keeping the position-error norm below 0.04 m on one real trajectory. The fixed paper does not publish the raw trace, repeat count, aggregation rule, confidence interval, exact component BOM, calibration record or failure log.
What did RCI validate? RCI fixed and hashed the HAL v3 and arXiv v1 paper files, checked item-level rights, traced fourteen observations and audited fourteen disclosure fields. It ran zero code, simulations, controllers, actuators or robots.
The paper supplies a promising author-reported transfer case across selected translational CDPR configurations. It does not validate arbitrary actuator counts, geometries, rotation, loads, cable physics, hardware stacks, reliability, safety or procurement suitability.
The policy is shared at actuator level
Instead of mapping a complete robot state directly to all motor commands, the reported policy acts on each actuator using normalized local fields. This design reduces dependence on a fixed actuator count, but architectural reuse is not the same as performance invariance.
| Field | Paper-reported value | RCI boundary |
|---|---|---|
| State | Current cable length, target cable length, t-3 cable length, motor current | Exact sensors and runtime interfaces are not identified |
| Control setting | 100 Hz | Not measured latency, jitter or deadline compliance |
| Nominal motor parameters | 6.8 A, 3420 rpm, 0.8 Nm, 0.123 Nm/A | No orderable motor or drive mapping |
| Drum radius | 0.0178 m | No exact winch, cable or encoder identity |
The disclosed transfer covers selected translational configurations
The training setup is a simulated four-cable CDPR with a point-mass platform and co-located cable attachments. The comparison uses four simulation seeds. The reported speed and target-position tests use ten episodes of ten seconds each.
For one four-cable comparison, the paper reports conventional-RL error spanning 0–0.5 m and actuator-level-policy error spanning 0–0.25 m. These are author-reported plot ranges, not RCI-recomputed metrics or universal error envelopes. The eight-cable transfer remains translational and point-mass based; rotational CDPRs are not validated.
The physical result is one bounded trajectory
The paper reports deployment on one eight-motor translational CDPR and shows one trajectory whose position-error norm stays below 0.04 m. Without the raw reference and measured trajectories, RCI cannot determine whether that threshold represents a maximum, observation range, filtered signal or another statistic.
The missing exact motor, drive, encoder, cable, drum, motion-capture, compute and power identities also prevent component-level comparison. A system-level demonstration cannot be converted into a motor, drive or computer recommendation.
A reusable experiment package is still missing
- Code and environment: no repository, dependency lock, versioned simulator environment or container was confirmed.
- Policy artifacts: no trained weights, reward configuration bundle or seed-level result package was found.
- Raw traces: no simulation trajectories or real reference/measured trajectory files are linked.
- Physical identity: the exact BOM, calibration, repeat count, acceptance criteria and failure log are absent.
- Compute identity: training hours are reported without CPU, GPU, memory, software-stack or energy context.
Fourteen disclosure checks
| Field | Status | Evidence | Boundary |
|---|---|---|---|
| fixed primary artifacts | complete | HAL v3 and arXiv v1 file identities, page counts, byte counts and audit-date SHA-256 values are fixed. | The SSRN PDF bytes were not independently compared. |
| item level file license | complete-with-restrictions | HAL v3 declares CC BY-NC-ND 4.0 for its PDF file. | No PDF redistribution, figure/table adaptation, extended-text reuse or commercial reuse of licensed expression. |
| public code | not-found | No repository link appears on the fixed item pages or in the reviewed paper. | Absence on audited surfaces does not prove no private code exists. |
| public raw data and weights | not-found | No raw simulation trajectories, real-robot traces, trained policy, weights or result bundle is linked. | Author-reported plots cannot be independently recomputed. |
| exact hardware bom | not-disclosed | The physical platform is described at system level. | Motor, drive, encoder, cable, winch, motion-capture, compute and power part numbers are missing. |
| real hardware protocol | incomplete | One physical trajectory and qualitative setup are reported. | Repeat count, trial acceptance criteria, calibration, environmental conditions and failure log are absent. |
| real hardware metric definition | incomplete | The paper reports a less-than-0.04 m position-error-norm trace. | No aggregation rule, confidence interval or downloadable values are supplied. |
| compute identity | not-disclosed | Training duration is reported. | CPU, GPU, memory, software stack and energy measurement are absent. |
| software environment | incomplete | OpenAI Gym is named. | No versioned environment, dependency lock, simulator version or reproducible container is supplied. |
| hardware replication | not-disclosed | One real CDPR is reported. | No independent operators, multiple specimens or external replication are reported. |
| arbitrary configuration claim | not-demonstrated | Selected four- and eight-cable translational configurations are reported. | Actuator counts, geometry, mass, rotation and cable physics are not exhausted. |
| safety validation | not-disclosed | A current-limit reward term is described. | Reward design is not functional-safety assessment or physical safety validation. |
| independent rci validation | not-performed | RCI inspected fixed paper artifacts and rights metadata. | RCI performed zero code, simulation, controller, actuator or robot runs. |
| component or procurement conclusion | out-of-scope | The paper evaluates a control approach. | It cannot rank motors, drives, cables, computers or CDPR platforms or establish procurement suitability. |
Method, work level and rights
RCI fixed the official Inria/HAL v3 and arXiv v1 artifacts, recorded file identities and rights metadata, and converted fourteen paper-reported observations into independently worded fields with explicit boundaries. RCI did not reproduce figures or tables, execute code, run a simulator, operate a controller, test an actuator or access the physical robot.
Highest completed work level: EV1 source verification. There is no EV3 recomputation, EV4 RCI physical test, EV5 external replication or EV6 accredited result. See the evidence work levels.
- All control and performance results remain author-reported; RCI executed no code or hardware.
- The disclosed evidence covers selected translational point-mass configurations, not arbitrary CDPRs.
- The real-hardware result is one reported eight-motor trajectory without a complete repeatable protocol or raw trace.
- No public code, dependency lock, environment, weights, raw results or exact hardware BOM was confirmed.
- The 100 Hz setting is not a measured latency or real-time guarantee.
- HAL's CC BY-NC-ND 4.0 file license does not authorize commercial reuse or adaptation of the paper's expression.
The official Inria/HAL v3 record declares a CC BY-NC-ND 4.0 file license. RCI uses independent factual wording and original analysis only; it does not mirror the paper, reproduce or adapt figures/tables, or reuse the paper's expression commercially. The arXiv v1 record is retained as a second fixed identity.
Download the evidence data
Download the immutable JSON release and record-level CSV. Stable aliases are current JSON and current CSV.
Suggested citation: Robot Component Index. “CDPR Actuator-Level Policy Evidence: 4-Cable Simulation to One 8-Motor Robot, Zero RCI Runs.” RCI 043, version 0.1.0, 2026-08-11. https://robotcomponentindex.com/research/cdpr-actuator-level-policy-evidence/
For version-aware BibTeX and machine citation records, use Cite RCI.