10 Robotics Breakthroughs of 2026, and What Each Does and Doesn't Show

Robotics news is unusually easy to over-read. A video of a robot folding a towel can mean a reliable product, a lab prototype or a single good take. This list takes ten 2026 developments that have a readable primary source and asks the same four questions of each: what was shown, when, who ran the demonstration, and what the evidence does not show.

What are the ten developments at a glance?

What happened with production humanoids?

1. Boston Dynamics unveils a production Atlas (5 January)

What was shown. At CES, Boston Dynamics unveiled the product version of its electric Atlas humanoid and said it would start manufacturing it at its Boston headquarters immediately. The release claims 56 degrees of freedom, a reach of 2.3 m (7.5 ft), a lift of up to 50 kg (110 lb) and an operating range of -20 to 40 degrees C. It says all 2026 deployments are already committed, with fleets going to Hyundai's training centre and to Google DeepMind, and that more customers are planned for early 2027. Hyundai Mobis will supply the actuators.

Who ran it. The company, in a keynote and a press release. What it does not show. The page gives no price, unit volume or performance benchmark, and the specifications have not been independently tested. The same release's boilerplate calls Atlas a platform 'currently in development', and the DeepMind partnership is described as aiming to add cognitive abilities, which is a goal, not a demonstrated feature.

2. The Atlas training centre opens in Georgia (21 September)

What was shown. Boston Dynamics said the first phase of the Robotics Metaplant Application Center at Hyundai's Metaplant America site near Savannah is operational. Atlas is being trained there on handling and sequencing automotive parts. The release says operations move to a building about ten times larger in 2027, component assembly follows by 2030, and Hyundai plans 25,000 units across its plants over the next few years.

Who ran it. Boston Dynamics. What it does not show. It reports no results or metrics from the centre so far. Heavy lifting and assembly are described as future work, and the 30,000-robot-a-year U.S. facility has no date. A training site is not a production line.

3. Agility Robotics unveils Digit 5 (15 September)

What was shown. Agility introduced Digit 5 as its first humanoid designed to work close to people without physical barriers. It lists a payload of up to 50 lb (22.7 kg), a 90-minute battery that charges in 9 minutes, swappable grippers and a body of 284 lb (129 kg) reaching 7.2 ft. It reports more than $300 million in multi-year Digit 5 orders as of May 2026, subject to contractual milestones, with early access expected in the first half of 2027 and general availability by the end of 2027.

Who ran it. The company. What it does not show. The safety design is described in the company's own words; the release offers no independent verification of the Digit 5 safety architecture. The orders are conditional, and the release is tied to a proposed merger with a special-purpose acquisition company, so it carries forward-looking-statement caveats. This is a reason to read it as an investor-facing announcement.

What are the big labs doing with robot models?

4. Gemini Robotics-ER 1.6 (14 April)

What was shown. Google DeepMind upgraded the reasoning model that plans for robots, with better pointing, counting, success detection and multi-camera reasoning, and a new ability to read gauges and sight glasses, developed with Boston Dynamics. It is available to developers through the Gemini API and Google AI Studio. DeepMind reports instrument-reading success of 93% with agentic vision, 86% without it, 67% for Gemini 3.0 Flash and 23% for the earlier ER 1.5.

Who ran it. DeepMind. What it does not show. These are vendor benchmarks. The page itself notes that agentic vision was on for instrument reading but unsupported by ER 1.5 and off for other tests, and that its single-view and multi-view results are not directly comparable. 'Safest model to date' is a claim relative to its own earlier models.

5. Gemini Robotics 2 (30 July)

What was shown. DeepMind announced three models: a vision-language-action model for humanoids and two-armed robots, a reasoning model that tracks multi-step tasks over minutes, and a smaller on-device model that it says can adapt to a new two-armed robot with a few hours of data. Demonstrations used an Apptronik Apollo 2 humanoid and other platforms. The reasoning model is on Google AI Studio; the other two are limited to early-access partners.

Who ran it. DeepMind; the page does not say who operated the demos. What it does not show. The company's own charts give success rates such as 68.4% for picking an object up from a table with one humanoid setup, and 36% for screwing in a bulb with another. Those figures show that multi-finger dexterity is still unreliable, and the team itself says dexterous manipulation 'remains challenging'. No methodology, latency or pricing is given.

6. NVIDIA's Astranomy report (14 October)

What was shown. An NVIDIA Research technical report studies a general-purpose AI model, named in the report as GPT-6 Astra, as a policy for language-instructed pick-and-place on a benchmark of 120 tasks. It tests three set-ups. The one where the model keeps a persistent working session reached 74.8% success, against 66.3% for direct action generation and 66.5% when the model supervises a robot foundation model. It also reports cost: about $36.40 per episode for the agent set-up against $1.96 for direct control.

Who ran it. NVIDIA researchers. What it does not show. The report says each task ran only five episodes, which is 'insufficient to characterize per-task variability', and that one comparison is not statistically significant. We read about the first 100,000 of 161,000 characters, so we may have missed its full limitations section. Speed and cost, not just success, limit this approach.

What do the research papers add?

7. Agent as Policy (11 September)

What was shown. A preprint reports a general-purpose AI agent that writes programs, issues motion commands and revises its actions from physical outcomes, with no task-specific training. Tasks included assembly from human videos, block building from goal images, dice flipping, targeted throwing and towel folding. The abstract claims success rates of at least 80% in seven of eight task configurations.

Who ran it. The authors, on their own set-up. What it does not show. The abstract does not say which configuration fell short or how far. It claims to beat earlier agentic robot systems without giving baseline numbers, and it is a preprint with revisions after its first posting.

8. Inductive generalisation for manipulation (19 June)

What was shown. The authors propose a reusable protocol for testing whether a manipulation policy generalises to progressively harder variations, and report that current vision-language-action models fail this test even when they handle simpler shifts.

Who ran it. The paper's authors. What it does not show. We read only the abstract, which gives no numbers. Its value is as a caution: a robot that looks general in a demo may be handling variations it has effectively seen before. Pair it with our guide to reading benchmark claims.

9. RoboDojo (5 July)

What was shown. A benchmark with 42 simulated and 18 real-world tasks, scored on generalisation, memory, precision, long-horizon execution and following open-vocabulary instructions. The authors integrated 30 policies and run a public leaderboard.

Who ran it. The paper's authors. What it does not show. The abstract states no findings about which policies do best. A benchmark also shapes what developers optimise for, which is why a public leaderboard needs scrutiny over time.

Did a humanoid really operate on animals?

10. Humanoid robots in surgery, in vivo feasibility (8 July)

What was shown. A preprint describes a humanoid-based laparoscopic teleoperation framework that uses general-purpose instruments. It was evaluated on a benchtop, in dry-laboratory user studies across surgeons of different experience, and in living pigs, and compared with established platforms such as da Vinci. The abstract names promise and 'key technical challenges that must be addressed before clinical deployment'. A search of Nature's site also lists a paper of the same title, but its page asked for a login, so we cite only the preprint.

Who ran it. The paper's authors; the page we read did not show their affiliations. What it does not show. A human surgeon controlled the robot remotely, so this is not autonomous surgery. The abstract reports no quantitative results, so we cannot tell you how well it performed. Animal studies are an early step, not evidence of safety in people.

What do the ten have in common?

• Demonstration is not deployment. A named customer, a training site and a committed order are different from a fleet doing useful work reliably.
• Success rates below 100% matter. Even a vendor's own charts show many tasks under 70%.
• Speed and cost are hidden constraints. The Astranomy report shows a large cost gap between approaches with similar success rates.
• Teleoperation is not autonomy. Ask who is controlling the robot, and for how much of the task.

For how vendor statements differ from verified findings, see vendor claims vs. verified findings. For the chips that run these models, see our AI hardware companies piece, and for what goes wrong when AI systems act in the world, read AI safety incidents and warning signs of 2026. Our methodology explains how we label evidence.

What are the limitations of this list?

• We read primary pages on 18 October 2026, mostly abstracts and announcements, not full papers. Findings in the papers may be more nuanced than their abstracts.
• Many well-known 2026 robotics stories are missing because we could not read a primary page for them. Absence does not mean they did not happen or were unimportant.
• Company dates and specifications can change. If you spot an error, tell us and we will record it in our corrections register.