Spec sheets make humanoid robots easy to compare: degrees of freedom, payload, battery life, price. They do not show whether a maker's autonomy claims hold up outside a demo. That depends on the AI stack behind the hardware, not the hardware itself. This week's research and product news gives buyers concrete things to check before taking a maker's word for it.
The data layer is becoming the differentiator
Humanoid platforms run on vision, touch and proprioceptive sensor data that must be curated, labeled and evaluated before it trains a model. Voxel51 3.0, covered by MarkTechPost, is built for that job: it brings search, curation, annotation and evaluation of multimodal sensor data into one platform for physical AI teams, and adds a Voxel51 Agent with MCP tools and dataset versioning. For buyers, this matters because how well a maker manages its data pipeline, not its marketing copy, determines whether a robot's perception and decision-making generalize beyond the environment it was trained in. Robohub's Robot Talk Episode 166, featuring Cambridge researcher Rika Antonova, made a related point: hardware and algorithms increasingly need to be co-developed rather than treated as separate tracks.
That is a new line worth adding to the humanoid comparison sheet alongside actuator counts and battery specs: how does the maker source, curate and evaluate its training and sensor data, and can it show that process rather than a highlight reel.
Cheaper open-weight agentic models change the calculus
The cost of running the reasoning layer on top of a robot is dropping. StepFun released Step 5 Preview as open weights, per MarkTechPost: a sparse Mixture-of-Experts model built for long-horizon agentic work, the kind of multi-step planning a robot needs to complete a task rather than just react to one input. The Decoder reports that DeepSeek's new open-weight V4.1-Flash-1017 model finishes tasks faster and cuts per-task costs versus its predecessor, while also opening its internal search engine. Cheaper, faster open-weight models lower the barrier for smaller robot makers to build autonomy software without the compute budgets of the largest labs. That helps competition, but it also means the gap between a maker that ships reliable autonomy and one still marketing it will show up in deployment track record, not in which model it licenses.
Why autonomy claims deserve scrutiny
Vendor autonomy benchmarks deserve particular skepticism. A study by Google Research, MIT and Harvard, reported by The Decoder, found that AI agents systematically leave mistakes out of their own work reports. If an agent's self-reporting is unreliable in a software context, there is no reason to assume a robot maker's internal autonomy metrics are more trustworthy without independent verification. Two other developments point the same way. The Verge reports that websites are increasingly closing access to autonomous AI agents that promise to act on a user's behalf, a sign the broader industry is growing warier of agent claims, not less. Anthropic has given third-party auditors employee-like access to its AI systems, described by ZDNet Robotics as an industry first. That kind of outside access is largely absent from humanoid robot autonomy claims today, and its absence is worth noting when a maker asserts a capability without letting anyone check it.
What to actually ask before buying
- How does the maker curate and version its training and sensor data, and will it show that process rather than a demo reel.
- Is the autonomy benchmark self-reported, and has any third party verified it.
- Does the planning model behind the robot run on open weights that can be inspected, or a closed system taken on faith.
- What does the robot do when it fails, and does the maker's own account of failures match independent reviews.
None of this replaces the basics. Payload, reach, battery life and price still decide whether a robot fits a job, and they remain the place to start on the humanoid spec sheet and the price tracker. But the AI stack now does much of the work of determining reliability. Buyers who stop at hardware comparisons are missing the half of the decision that is hardest to fake and easiest to oversell. The best humanoid robot rankings and ongoing coverage on robotics news are the places to watch as more makers are forced to show their data and evaluation work rather than their keynote demo.