The direct answer
Which robot components in RCI now have independent evidence? Catalog v0.4.1 contains bounded independent records for the ROBOTIS XM540-W270-R, RealSense D455, Livox Mid-70, NVIDIA Jetson Orin Nano 8GB and Molicel P42A. The D455 has two separate independent findings, so five component records contain six V-state claims.
Does a V state mean the product is independently verified? No. RCI assigns V to one claim, not to a product, company or specification sheet. A study that deploys an actuator can establish exact-model use without validating its torque rating. A mapping experiment can report a LiDAR-and-algorithm result without validating the sensor’s maximum range. The boundary is part of the evidence.
The component count remains 75. Catalog v0.4.1 adds five evidence records to the 544-claim v0.4.0 baseline, producing 549 claims: M 525, V 6, D 1, C 3 and U 14. No manufacturer claim was silently promoted.
The six records at a glance
| Exact component | What the independent source reports | What it does not establish |
|---|---|---|
| ROBOTIS DYNAMIXEL XM540-W270-R | Exact model used as the rotation motor in a weaving-gripper experiment repeated five times. | Deployed-use evidence; no independent torque, backlash, thermal or life rating. |
| RealSense Depth Camera D455 | Two studies: distance/inclination-dependent depth behavior, plus a 2.47 mm mean P_S error in a close-range ISO procedure. | Two test setups and selected metrics; not every unit or the full 6 m ideal range. |
| Livox Mid-70 | 1.8 cm static and 4.3 cm dynamic dimensional registration error with the paper’s proposed method. | Sensor-plus-algorithm mapping result; not intrinsic range precision or maximum range. |
| NVIDIA Jetson Orin Nano 8GB | 31.8 J for CUDA versus 28.3 J for SYCL across four selected AI benchmarks. | 15 W, JetPack 5.1.2 and specific implementations; not Super-mode TOPS or robot-system power. |
| Molicel INR-21700-P42A | 45 cells screened; 12 selected cells characterized at 5°C, 25°C and 40°C. | Dataset/test-scope evidence; not a 45 A validation, pack test or population certificate. |
XM540-W270-R: independently documented use, not a torque certificate
The 2023 Nature Communications paper “Grasping through dynamic weaving with entangled closed loops” identifies the exact XM540-W270-R in its methods. The actuator rotated the study’s PET-strip gripper. The experiment moved the gripper from an initial −70° angle to four final-angle conditions, applied a vertical pull until failure and repeated experiments five times. A load cell and a torque sensor instrumented the apparatus.
RCI therefore publishes one V-state deployed-experiment claim. The paper does not independently measure the servo’s 10.6 N·m manufacturer stall-torque figure, 0.25° backlash claim, continuous thermal output or service life. Those fields keep their existing M or U states.
D455: two studies, two different questions
The RealSense D455 now has two independent claims because two papers test different behaviors. A 2023 ISPRS study, “Geometric Accuracy Investigations of Intel RealSense Depth Cameras”, found that depth noise increased with distance and that accuracy degraded sharply at high surface inclinations under its checkerboard and indoor-mapping setups. That is a directional behavior, not a universal error percentage.
A separate 2021 Sensors paper, “Metrological Characterization and Comparison of D415, D455, L515 RealSense Devices in the Close Range”, applied an ISO 10360-13 procedure over 500–1500 mm indoors under controlled brightness. For the tested D455, the paper reports a mean PS error of 2.47 mm with a 1.42 mm standard deviation.
The second result is deliberately not restated as “2.47 mm D455 accuracy.” It describes one metric, one tested unit and one close-range setup. It neither validates the manufacturer’s full 0.6–6 m ideal-range statement nor predicts unit-to-unit production variation.
Mid-70: a mapping-system result, not intrinsic sensor precision
The 2024 Sensors paper “Non-Repetitive Scanning LiDAR Sensor for Robust 3D Point Cloud Registration in Localization and Mapping Applications” evaluates a Livox Mid-70 with a proposed registration method. The static experiment used 35 scans in a room; the dynamic experiment used 357 scans from a mobile platform. Both used 0.2 s integration. The sensor was warmed for more than 30 minutes, measurements below 1 m were discarded and physical room dimensions supplied the reference.
The proposed method produced average dimensional registration errors of 0.018 m in the static experiment and 0.043 m in the dynamic experiment. Those numbers belong to the entire disclosed acquisition-and-registration pipeline. They are not substitutes for the manufacturer’s intrinsic range-precision condition, and they do not establish a detection range.
Orin Nano 8GB: energy changes with code and workload
The 2024 Journal of Supercomputing paper “SYCL in the edge: performance and energy evaluation for heterogeneous acceleration” identifies an 8 GB Jetson Orin Nano board. It ran at 15 W with JetPack 5.1.2; tegrastats measured package power. Across four selected HeCBench AI tests, the reported total was 31.8 J for CUDA and 28.3 J for SYCL, a 12% difference in that experiment.
This is valuable because it shows that programming model and workload can materially change energy use on the same board. It does not verify 67 sparse INT8 TOPS, JetPack 6.2 Super mode, peripheral consumption or whole-robot battery draw. The paper itself warns that energy behavior varies with the application and tuning.
P42A: a public multi-cell, multi-temperature characterization dataset
The 2025 Scientific Data paper “High-power lithium-ion battery characterization dataset for stochastic battery modeling” is based on exact-model Molicel INR-21700-P42A cells. Phase 1 screened 45 cells at 25°C using capacity and pulse-resistance measurements. Twelve cells were then selected from the observed distributions.
Phase 2 characterized those 12 cells at 5°C, 25°C and 40°C using constant-discharge tests, high-C-rate pulse/GITT tests and electrochemical impedance spectroscopy. RCI records the population and protocol scope as V. It does not convert the dataset into independent certification of the manufacturer’s 45 A continuous-discharge limit, a battery-pack result or a guarantee about every P42A cell.
What changed—and what V still means
The previous catalog had one V-state claim. This release adds five more without changing any M-state value. That separation matters: independent evidence should supplement a product record, not overwrite a different metric merely because both contain a familiar unit.
RCI uses V when an external record directly supports the exact model and stated result under a disclosed boundary. The minimum publication test is:
- Identity: the source identifies the same model or orderable variant; family-level ambiguity is disclosed rather than hidden.
- Observation: the source reports an experiment, deployed bill of materials, certification record or other direct external result—not a copied vendor table.
- Condition: the apparatus, sample, workload, environment or software state needed to interpret the result remains attached.
- Non-expansion: the claim states what the source did not test, especially when a nearby manufacturer field looks superficially similar.
A V claim is stronger evidence for its exact sentence. It is not a product score. Buyers still need application-specific validation for safety, integration, reliability and production acceptance.
Download the claims and inspect the source trail
Download the immutable v0.4.1 JSON catalog or claim-level CSV. Each row includes its evidence state, source title and URL, source location, condition, editorial note, verification date and reviewer. Stable aliases remain available at /data/components.json and /data/components.csv.
Suggested citation: Robot Component Index, “Five Robot Components, Six Independent Evidence Records — and Their Limits,” RCI 013, dataset v0.4.1, 2026-08-10.