WuJi Hand 2 vs SharpaWave: what RLWRLD’s comparison establishes

RLWRLD puts a purchasing question at the centre of its robot-hand comparison: what does a higher price buy? In the description of How we actually test a robot hand, the company says it tested WuJi Hand 2 and SharpaWave side by side, listing them at $19K and $50K respectively. Those figures come from RLWRLD; the supplied material does not specify the currency beyond the dollar symbol, comparable configurations or included services. RLWRLD video

The comparison’s value depends on connecting that price framing to documented behaviour. RLWRLD names durability, backdrivability, heat management, repeatability, “palm/min grasp” and control under teleoperation as testing categories. These identify the questions the company says it investigated. They do not establish how either hand performed.

This analysis draws on the supplied video metadata and description, technical-blog excerpt, report abstract and repository excerpt. The footage, scoreboard and underlying test records were not reviewed. The supplied upload date is September 3, 2026; it does not establish when testing occurred. A chapter marker places the scoreboard at 09:19, but no scores or scoring rules are included in the available evidence. A comparative winner therefore cannot be established here.

To make the announced categories useful, each needs a defined procedure. A durability conclusion would need operating exposure, cycle counts, failure criteria and maintenance records. Heat-management findings would need temperatures, measurement locations and operating duration. Repeatability would need repeated attempts and the variation between them. These are requirements for assessing a conclusion, not assertions about tests RLWRLD did or did not conduct.

The phrase “palm/min grasp” also needs clarification. The supplied description does not define it, so it cannot responsibly be converted into a specific grasp metric. Backdrivability likewise needs a stated measurement method and operating condition before readers can compare results.

Control mode is another essential part of the comparison. The description explicitly includes teleoperation. An operator-controlled action can inform an assessment of that hand-and-control setup, but it does not by itself establish that an autonomous model can select and execute the same action. Operator experience, interface and calibration would all need to be documented to interpret a comparison.

The chapter list places an RLDX-1 segment at 14:57, after the scoreboard. That structure does not identify which earlier sequences, if any, used the model. Crediting RLDX-1 with the hand-testing outcomes would go beyond the supplied evidence. Video description and chapter list

RLWRLD’s supporting material explains why it connects hardware testing with model development. Its technical blog describes RLDX-1 as a model focused on dexterous manipulation and says the company has operated more than ten five-finger hands in-house. The blog and repository describe an architecture incorporating motion information, memory and physical sensing, including tactile and torque signals. These are the company’s descriptions of its system; the excerpts do not establish which inputs or features were active in this comparison. RLDX-1 technical blog, RLDX-1 repository

The claim that RLDX-1 works across many hands also needs a precise reading. Running on the company’s hardware does not establish equal performance across those hands, support for every commercial model or transfer to an unfamiliar hand without additional training or calibration.

The supporting evaluations answer separate questions. The technical-report abstract describes author-reported results on ALLEX humanoid tasks. The repository presents simulation results for models fine-tuned on each benchmark’s training set. Neither supplies the missing measurements of wear, heat or repeatability for WuJi Hand 2 and SharpaWave. Technical-report abstract, Repository benchmark description

The evidence supports an account of RLWRLD’s announced testing scope and its rationale for working with multiple hands. It does not establish which hand performed better or whether either offers better value. The next useful evidence would connect the scoreboard to test definitions, trial counts, failures, control modes and comparable price quotations. No independent evaluation is included in the supplied material.