2026 Humanoid Robot Benchmark: Figure 03, Optimus Gen-3, Unitree G1, Digit

Every humanoid spec sheet is a marketing document, and the numbers on it are not comparable. One vendor quotes peak payload, another quotes sustained payload, a third quotes nothing at all. Battery runtimes are given “at idle”, “at typical load”, or “at peak performance”, and the three can differ by a factor of three. A humanoid robot benchmark that simply lines those figures up in a table produces a confident-looking comparison of things that were never measured the same way.

That matters now because the buying decisions are real. Agility Robotics unveiled Digit 5 in September 2026, Unitree listed on the Shanghai exchange in August, Figure is building robots at a rate of roughly one per hour, and Tesla has still not shown Optimus V3. This guide compares the four platforms on what is actually verifiable, flags every number that is not, and replaces the usual spec table with a five-axis evidence framework. You will leave knowing which claims to trust, how to turn a battery figure into a cost per productive hour, and what to demand from a vendor before a pilot.

What this covers: the shared architecture of a humanoid stack, a verification-tier method, platform-by-platform facts as of 30 October 2026, battery and cost arithmetic, failure modes, and a pilot checklist.

Context and Background

In 2024 the question was whether a bipedal robot could do useful work at all. By late 2026 that has been narrowly answered: yes, in structured, repetitive tasks such as moving totes, sequencing parts, and sorting sheet-metal components. The open question has shifted to economics and reliability, which no vendor demo can settle.

Four platforms define the conversation, and they are not really competitors in the same market. Figure AI (San Jose) builds a general-purpose humanoid with an in-house vision-language-action model, Helix, and its own factory, BotQ. Tesla is building Optimus for its own plants first. Unitree (Hangzhou) sells inexpensive, highly mobile research humanoids and has just become, per Bloomberg-sourced reporting, the first Chinese humanoid maker to list on a mainland exchange. Agility Robotics (Oregon) sells Digit as a robots-as-a-service product for logistics and has designed the new Digit 5 around industrial safety evaluation.

The contrast with industrial automation is instructive. A six-axis arm ships with a datasheet defined by ISO 9283 performance criteria and a payload curve at stated reach. Humanoids have no equivalent. ISO 25785-1, a safety standard for dynamically stable industrial mobile robots, is under development, and the safety conversation currently leans on adapting ISO 10218 and ISO 13482 concepts. The IEEE has a humanoid study group as well. Until measurement standards land, the burden of comparison falls on the buyer.

If you are building the software side, the middleware most of these robots interoperate with in research settings is ROS 2; our ROS 2 Jazzy Jalisco migration guide covers the current long-term-support release. For safety background, NVIDIA’s Isaac and Halos platform pages describe the stack Agility named as its safety launch partner.

A note on provenance. An earlier version of this post, published in April 2026, carried a spec table that did not hold up against vendor and independent sources: it listed the Unitree G1 at 170 cm and 55 kg when it is 132 cm and about 35 kg, and quoted battery capacities in the 7 to 12 kWh range that no tracked source supports. This rewrite starts again from sources and discards anything it cannot trace.

All four robots follow the same logical stack, even though the components differ in size and maturity. Understanding it explains why a single “benchmark score” is the wrong abstraction.

Direct answer: A modern humanoid runs a slow vision-language model for task reasoning, a fast visuomotor policy for manipulation, and a whole-body controller that turns intent into joint torques, wrapped in a safety layer. Performance depends less on any one layer than on how well the layers hand off to each other under latency and energy limits.

Figure 1: The common humanoid control stack. Perception feeds a slow reasoning layer and a fast policy; the whole-body controller closes the loop through joint feedback while a safety monitor can cut actuator power.

The diagram shows a perception block feeding two policy tiers, a controller, actuators, and a feedback path through the IMU and joint encoders. The safety monitor sits beside the controller with authority over the actuators; this separation is what an OSHA-style evaluation interrogates.

Layered policies: slow reasoning, fast action

Figure has published the clearest public description of this split. Helix, introduced in 2025, used a “System 2” vision-language model of about 7 billion parameters running at 7 to 9 Hz, and a “System 1” visuomotor transformer of about 80 million parameters running at 200 Hz. Helix 02, released 27 January 2026, added a “System 0” of roughly 10 million parameters at 1 kHz for whole-body balance and contact control, according to the secondary source tracked for this post. The numbers are Figure’s, and I have not seen independent replication.

The point of the split is bandwidth. A 7B model cannot emit joint targets at 200 Hz on an embedded GPU, but fingers closing on a tote need exactly that rate. So the large model emits a latent goal slowly and the small model interpolates it into motion. Unitree has taken a related route with its open UnifoLM models, whose VLA variant was announced in January 2026.

The older version of this post asserted parameter counts for Optimus and Digit policies, and latencies such as “5 ms IK for Figure”. None of those is published. Treat any such per-platform inference number as folklore until a vendor documents it.

Whole-body control and the safety layer

Below the policies sits the whole-body controller. For a bipedal robot, its job is to hold balance while the arms move and carry load, which means the center of mass shifts every time a 20 kg tote is lifted. Controllers use a mix of model-predictive control and learned policies trained in simulation with reinforcement learning, then transferred to hardware. Figure said in April 2026 that a balance policy called Vulcan tolerates the loss of up to three lower-body actuators, an unusually specific robustness claim worth testing in a pilot.

The safety layer is where Agility has staked a differentiated claim. Digit 5 uses cameras, depth sensors and inertial units to detect people, slow, stop, or sit down and cut motor power, with an emergency stop and safety display on the rear. Agility says it is the first humanoid designed for field evaluation against OSHA industrial safety standards, with NVIDIA’s IGX Thor and Halos for Robotics as the first safety-stack partner. That is a vendor claim about design intent, not a certification.

Why the hands decide the use case

Locomotion gets the demos, but hands determine which jobs are feasible. Figure says each Figure 03 hand has 16 degrees of freedom with fingertip tactile sensing that detects forces as small as 3 grams. The Unitree G1 spans a wide range: the base model has a simple gripper, while the EDU Ultimate variants fit Dex3-1 three-finger hands (7 degrees of freedom) or Dex5-1 five-finger hands (20 degrees of freedom with up to 94 tactile sensors, per a secondary spec source). Digit moves totes with swappable grippers, including a claw-like option for large bins, which is a deliberate choice: a tote-handler does not need fingers.

Tesla’s hand is the most speculative. An engineer-sourced report claims 22 degrees of freedom for the Gen 3 hand, while another analysis says a previously patented hand design was scrapped in April 2026. Tesla itself has published no official dimensions, weight, payload, or runtime.

How to Read Humanoid Claims: Five Axes and Three Tiers

My thesis is that a humanoid robot benchmark should not rank robots. It should rank claims, by how they were verified, and then convert only the verified ones into a cost per productive hour. This is more useful than a spec table because the decision a buyer actually faces is not “which robot has the highest payload” but “which vendor’s claims survive contact with my floor”.

Five axes that matter

Verified deployment tier. Has anyone other than the vendor seen the robot do paid work?
Energy duty cycle. Runtime multiplied by charging or swap logistics, expressed as the fraction of a shift the robot is productive.
Payload at reach, sustained. Not peak, not momentary, and not at the elbow.
Autonomy ratio. The fraction of tasks completed without human intervention, plus mean time between interventions.
Cost per productive hour. Lease or amortized price divided by hours of useful work, including supervision.

None of the vendors publishes axes 4 and 5 in a form you can audit. That is the gap an honest comparison has to expose rather than paper over.

The verification tiers

I borrow the tiering used by independent trackers such as the Humanoid Reality Index: a platform is “deployed” only when independent sources show repeated paid work, “pilot” when a named customer ran a trial, and “demo” when there is only a public showing or vendor claim.

Figure 2: Claim-verification flow. A spec or deployment number is only as strong as the weakest question it fails: measured, at a named customer, paid and repeated.

The flow asks three questions in order. A number that was never measured is a demo-tier claim no matter how precise it looks. A number measured but not at a named customer site is a vendor trial. Only paid, repeated work at a named customer reaches the deployed tier.

Applying this to October 2026 evidence: Agility Digit sits at the deployed tier, with independent reporting of work at GXO, Schaeffler, Toyota Motor Manufacturing Canada and pilots at Amazon. Figure sits between pilot and deployed, with a verified, vendor-documented 11-month BMW Spartanburg run for the earlier Figure 02 and a still-thin account for Figure 03. Unitree is a volume seller with strong shipment evidence but mostly research and education customers. Optimus sits at the demo tier.

A worked scoring example

Take one claim from each vendor and ask the questions in Figure 2. “Digit has moved over 100,000 totes at a GXO site” passes: it is measured, named, and reported by independent outlets, although the ledger I consulted lists it as both a pilot and a deployment, which tells you the category boundary is fuzzy. “Optimus will reach 1 million units per year at Fremont” fails the first question because it is a plan. “Figure 03 works for BMW at about $25 per robot-hour” fails because a single trade outlet reported it without sourcing and neither company has published the contract. “Unitree has shipped X thousand G1s” is a volume claim I could not confirm from a primary source in this research, so I leave the number out.

Platform Facts as of 30 October 2026

What follows is what I could trace to a source. Where a number came from a secondary aggregator rather than the vendor, it is marked. Always re-verify against the vendor datasheet before a purchase order.

Figure 03

Figure unveiled the third-generation robot on 9 October 2025. The product page gives a height of 5 ft 8 in (about 173 cm), though TIME’s cover story said 5 ft 6 in, and a weight of about 61 kg, roughly 9 percent lighter than Figure 02. The rated payload is 20 kg.

The battery is the most important correction to common lore. The F.03 pack is integrated into the torso at about 2.3 kWh, with roughly five hours of runtime at peak performance, per Figure’s battery documentation as summarized by a secondary source. Charging is wireless and inductive through coils in the feet, at about 2 kW with active cooling. Figure’s launch material also describes a palm camera in each hand. Spec directories list 44 degrees of freedom and eight cameras, but Figure has not published those, so I treat them as unverified.

Manufacturing is where Figure has the strongest recent evidence. On 29 April 2026 the company said it had delivered more than 350 units, raised its rate from one robot per day to one per hour, and reached end-of-line first-pass yield above 80 percent with a 99.3 percent yield on the battery line. The BotQ facility in San Jose has an initial capacity of up to 12,000 robots per year and a stated goal of 100,000 over four years. Those are capacity claims, not shipments.

On deployment, the documented result is for Figure 02: an 11-month deployment at BMW’s Spartanburg plant, with more than 1,250 operational hours, over 90,000 sheet-metal parts handled, and contribution to more than 30,000 X3 vehicles. For Figure 03, BMW’s own 25 June 2026 statement describes a parts-sequencing logistics use case. A separate trade report of a 40-robot paid deployment is single-sourced and I would not rely on it.

Tesla Optimus

There is no Optimus benchmark because there is no published Optimus specification. As of the most recent tracking I found (August 2026), Tesla had not revealed V3, had not published dimensions, weight, payload, runtime, speed or ingress rating, and had given no price or preorder. Figures such as “180 cm, 45 kg, 20 km/h, IP68” circulate online and are unsupported.

What is on record is timing. Elon Musk first pointed to a Q1 2026 reveal, then “probably middle of this year”. Tesla’s 22 July 2026 update said first-generation Optimus production lines were being installed at Fremont with production “anticipated later this year”, and Musk described the ramp as “quite flat and long”. Reports of a late-July reveal traced to unofficial sources. Musk’s stated target for public sales is the end of 2027, given at Davos in January 2026. His earlier projection of about 10,000 units in 2025 did not materialize.

Capacity ambitions of 1 million units per year at Fremont and 10 million at Giga Texas appear in secondary coverage without a primary Tesla source in what I reviewed, so treat them as aspirations. I found no confirmed useful work by Optimus units in Tesla factories. If something has changed since mid-August, I could not locate it, and the reader should check Tesla’s investor materials directly.

Unitree G1 (and its siblings)

The G1 is the cheap, small, athletic platform, and it is routinely confused with larger humanoids. It is 132 cm tall and about 35 kg, folds to about 690 mm for transport, and walks at up to about 2 m/s. Unitree unveiled it at ICRA 2024 in May and began production in August 2024.

Variants matter more than any single spec. The base G1 has 23 degrees of freedom and a simple gripper. EDU Plus reaches 29, and the EDU Ultimate versions go to 41 or 43 with dexterous hands. Maximum knee torque is quoted at 90 N·m for the standard and 120 N·m for EDU. Arm payload is small: about 2 kg standard and about 3 kg for EDU, according to the secondary source. That is a research-platform figure, not a logistics figure.

Prices are a range, not a number. Reporting at the time of the August 2026 IPO listed the G1 from $13,500; a spec aggregator lists the base at $16,000 to $17,990 and EDU variants from about $43,900 to $73,900. The sensor suite includes a Livox MID-360 lidar and an Intel RealSense depth camera, with EDU variants using an NVIDIA Jetson Orin at up to 100 TOPS. The battery is a 13-series pack of 9,000 mAh, and runtime is around two hours of mixed use.

Unitree’s larger family includes the H2 (from $29,900, about 182 cm, about 70 kg, 31 degrees of freedom) and the consumer-priced R1 (from about $5,900). Unitree debuted on the Shanghai Stock Exchange on 18 August 2026, with reports of a first-day gain of 460 percent. For the first nine months of 2025, 73.6 percent of its humanoid revenue came from scientific and educational institutions and about 9 percent from industrial use. That customer mix is the honest summary: Unitree is democratizing research humanoids, not yet staffing factories.

Agility Digit 5

Digit is the only one of the four with a clear commercial track record in logistics, and the September 2026 refresh targets its known weaknesses. Agility unveiled Digit 5 on 15 September 2026. It stands 5 ft 11 in, weighs 284 lb (about 129 kg), and extends its reach from 5.5 ft to 7.2 ft. It lifts 50 lb (about 23 kg) repeatedly, a 40 percent increase over the prior model, and uses about 30 motors with motor-level control at roughly 1,000 Hz.

The battery story is a deliberate trade. Digit 5 offers about 90 minutes of operation and recharges in about 9 minutes, which would support more than 20 hours of work per day if the robot reliably returns to its dock. The previous-generation figures I found (35 lb, around 4 hours) show that Agility has chosen shorter runtime with much faster charging.

Availability is early access in the first half of 2027, with wider release later that year. Production is planned at the 70,000 sq ft RoboFab in Salem, Oregon, with capacity for up to 10,000 robots per year. Agility reported more than $300 million in conditional commitments as of May 2026. Pricing is unpublished; a June 2026 SPAC investor presentation footnote assumed $8,500 per month per Digit as SaaS pricing. That is an assumption in a deal document, not a rate card.

Deployment evidence is the strongest of the four: independent reporting of work at GXO in Georgia (more than 100,000 totes), at Schaeffler, at Toyota Motor Manufacturing Canada under a robots-as-a-service agreement, and trials at Amazon. Several headline numbers, such as 65,000 operating hours, have no named customer behind them.

Side-by-Side Table: What Is Verified

Dimension	Figure 03	Optimus (V3)	Unitree G1	Agility Digit 5
Height	~173 cm (vendor)	Not published	132 cm	5 ft 11 in
Mass	~61 kg	Not published	~35 kg	284 lb (~129 kg)
Payload	20 kg rated	Not published	~2 to 3 kg arm	50 lb repeated
Battery	~2.3 kWh (secondary)	Not published	~0.42 kWh (derived)	Not published
Runtime	~5 h at peak	Not published	~2 h mixed	~90 min
Charge	2 kW wireless	Not published	~1.5 to 2 h	~9 min
Hands	16 DoF per hand	22 DoF (reported)	Gripper to 5-finger	Swappable grippers
Price	Unpublished	Unpublished	From ~$13.5k to ~$74k	Unpublished, ~$8.5k per month assumed
Evidence tier	Pilot to deployed	Demo	Volume seller, research	Deployed

The G1 battery energy is my arithmetic: a 13-series lithium pack has a nominal voltage near 46.8 V, and 46.8 V times 9 Ah is about 421 Wh. The table makes one point clearly: three of the four columns have blanks that the old post filled with numbers nobody can source.

Deeper Analysis: Energy, Duty Cycle and Cost per Productive Hour

Specs describe a robot. Economics describes a deployment. The bridge is the energy duty cycle, and it is where product philosophies diverge most visibly.

Figure 3: From pack energy to cost per productive hour. Task power draw sets runtime, charging or swapping sets downtime, and interventions reduce productive time further.

The diagram chains five quantities. Usable energy divided by average task power gives runtime. Adding charge or swap time gives a duty cycle. Uptime and interventions reduce that further, and only then does a price turn into cost per productive hour.

Reading the battery numbers

Start with implied average power, which is energy divided by runtime. Using the sourced figures: Figure 03’s 2.3 kWh over about five hours at peak performance implies roughly 460 W. The G1’s derived 0.42 kWh over about two hours implies roughly 210 W. Both are plausible magnitudes for a walking, working robot of their size, which is a useful sanity check. By contrast, the old post’s “12 kWh, 10 to 14 hours” implied about 1 kW for a 70 kg robot, and the cited battery sizes were unsourced.

For Digit 5 the pack energy is not published, but the shape of the design is: 90 minutes on, 9 minutes charging. The duty cycle is 90 divided by 99, or about 91 percent, ignoring docking travel. That matches the claim of more than 20 hours of work per 24-hour day only if no queueing for the charger occurs. With one charger per robot, it holds. With a shared charger and ten robots, it does not.

Figure takes the opposite design stance: a larger runtime window at peak, plus wireless charging through the feet so the robot can recharge opportunistically while standing at a workstation. Whether that yields a high duty cycle depends on whether the workflow includes natural pauses, which is a floor-layout question rather than a robot question.

Cost per productive hour, with stated assumptions

Here is an illustrative calculation, not a measurement. Take the $8,500 monthly SaaS price that appears in the Agility SPAC footnote as an assumption. If the robot works 20 productive hours a day for 30 days, that is 600 hours, or about $14 per hour. If it works one 8-hour shift on 22 days, that is 176 hours, or about $48 per hour. The same contract price differs by a factor of 3.4 depending on utilization, which is why vendors emphasize 20-hour days.

Now add supervision. If one human supervisor oversees ten robots and costs $40 per hour fully loaded, that is $4 per robot-hour. If interventions double that ratio to one supervisor per five robots, it is $8. These numbers are my assumptions, and the point is structural: supervision is a second-order cost that is small in a good deployment and dominant in a bad one.

Compare with the single reported Figure figure of about $25 per robot-hour. It is unverified, but it falls inside the same order of magnitude as the Digit arithmetic, which suggests early commercial pricing is converging on tens of dollars per hour. A human warehouse worker’s loaded cost varies widely by region, so I will not claim a break-even. The correct output of this exercise is a template you fill in with your own utilization and wage data.

What payload numbers hide

A 20 kg rating on Figure 03 and a 50 lb (22.7 kg) repeated lift on Digit 5 look close, but the geometry differs. Digit’s reach grew to 7.2 ft so it can place totes on higher shelves, and payload at full extension is the number that matters, not payload close to the body. Neither vendor publishes a payload-versus-reach curve.

The G1’s 2 to 3 kg arm payload is not a deficiency but a different product class. It is a platform for researchers developing manipulation policies, where 43 degrees of freedom and open interfaces matter more than lifting.

Autonomy ratio: the missing number

The most decision-relevant metric is how often a human has to step in. A robot with a 99 percent per-pick success rate sounds excellent until you compute the interruptions: at 600 picks per hour, that is six failures every hour, each requiring recovery or remote assistance. Nobody publishes this. When a vendor says “100,000 totes moved”, the follow-up questions are over what time, with how many robots, and how many assists.

Figure 4: A typical logistics task loop. The fleet cloud assigns work, the safety layer can interrupt the robot when a person is nearby, and failures are logged and routed to remote assistance before the robot resumes.

The sequence shows why intervention counts are not a single number. Some interruptions are safety stops that are working correctly, some are grasp failures, and some are fleet-software issues. A useful vendor report separates them.

Trade-offs, Gotchas, and What Goes Wrong

Vendor-defined payload. “Peak”, “rated”, “sustained” and “momentary” are not interchangeable. The old post quoted a 30 kg momentary Figure payload that no source supports; the sourced figure is a 20 kg rating. Ask for payload at a specified reach and duration, with a thermal soak, because actuators derate when hot.

Battery safety and certification. Figure said it has built a fireproof battery pack, and a robot that charges wirelessly through its feet adds a high-power inductive link near people. Batteries in torso-integrated packs complicate swap logistics and thermal management. Digit’s design trade, short runtime and fast charge, concentrates risk on charger availability instead.

Safety certification gaps. Dynamically stable robots do not stop safely by cutting power: they fall. That is why Agility’s “sit down and shut off the motors” behavior and Figure’s actuator-loss tolerance are interesting, and why no humanoid yet has the kind of functional-safety certification a collaborative arm carries. Do not allow humanoids to share space with untrained workers on the basis of a vendor brochure.

Teleoperation dependence. Demo videos often hide human operators. Training data for policies such as Helix comes from teleoperation, and in deployment a remote-assist path may carry real workload. Ask what fraction of completed tasks involved a human in the loop, and whether that human was onsite or remote.

Lock-in through fleet software. Robots-as-a-service pricing bundles hardware, software and support. That lowers capital risk but means your process data and workflow definitions live in the vendor’s cloud. Negotiate data export and exit terms before the pilot, not after.

Corporate and market risk. Agility is going public through a SPAC merger, Unitree’s stock moved more than 460 percent on its first day, and Tesla’s timelines have slipped repeatedly. A vendor’s financing story is a procurement risk. None of this is investment advice.

Counterfeit precision. Specification directories and AI-generated articles often fill gaps with plausible numbers. The 44 degrees of freedom and eight cameras listed for Figure 03 and the “70B-parameter Optimus policy” circulating in older writeups are examples of precision without provenance.

Practical Recommendations

Pick the platform by the job, then insist on evidence for that job.

For tote handling and logistics, Digit is the only one of the four with independent deployment history, and Digit 5 targets the right problems: payload, reach, charging and a safety-evaluation path. The catch is timing, because early access is not until the first half of 2027. Ask whether current-generation Digit units are enough for a pilot now.

For automotive and general manufacturing, Figure has the most credible manufacturing-scale story, and the BMW Figure 02 run is a real data point. Treat the Figure 03 BMW details as unverified until published by BMW or Figure.

For research and education, the Unitree G1 is the pragmatic choice. It is inexpensive, has open interfaces, and offers variants up to 43 degrees of freedom. Do not buy it expecting to carry 10 kg.

For Optimus, there is nothing to buy or evaluate. Watch for a reveal with an actual spec sheet, and reserve judgment.

A pilot checklist that works across vendors:

[ ] Require a written definition of payload, runtime and speed, including conditions and reach.
[ ] Measure intervention rate and mean time between interventions on your own task for at least two weeks.
[ ] Separate safety stops from task failures in all reporting.
[ ] Compute cost per productive hour with your utilization, not the vendor’s.
[ ] Confirm charger count and layout against fleet size.
[ ] Get data export and exit rights in writing.
[ ] Ask for the safety case: standards referenced, risk assessment, and who signs off.
[ ] Ask for two named reference customers whose sites you can call.

If you also run digital-twin or simulation workflows for layout planning, model charging queues explicitly. A twin that ignores the charger will overestimate throughput for every platform here, especially Digit’s fast-charge design.

Frequently Asked Questions

Which humanoid robot is best in 2026?

None is best overall, because the four target different jobs. Agility Digit has the strongest independent deployment record in logistics, Figure has the most credible manufacturing scale-up, Unitree G1 offers the lowest-cost research access, and Tesla Optimus has no published specifications. The right choice depends on your task, your floor layout, and the evidence a vendor can show for that specific task, not on a headline spec.

How much does a humanoid robot cost in 2026?

It ranges from under $6,000 for Unitree’s consumer-oriented R1 to roughly $13,500 to $74,000 for G1 variants and about $29,900 for the H2. Figure, Tesla and Agility do not publish prices. A June 2026 investor presentation assumed $8,500 per month for Digit under software-as-a-service pricing, and one unverified report cited about $25 per robot-hour for Figure. Treat these as indications only.

Is Tesla Optimus available to buy?

No. As of the latest tracking I found, Tesla had not revealed Optimus V3, published specifications, or opened orders. Tesla said first-generation production lines were being installed at Fremont with production anticipated later in 2026, and Musk has stated a goal of public sales by the end of 2027. Those are targets that have slipped before, so plan as if no purchase is possible before 2028.

How long does a humanoid robot battery last?

Far shorter than a human shift. Figure 03 quotes roughly five hours at peak performance from a pack of about 2.3 kWh, the Unitree G1 manages around two hours of mixed use, and Digit 5 is rated at about 90 minutes before a roughly nine-minute recharge. Real runtime depends on payload, walking time and temperature. Ask for runtime measured on your task, then compute the duty cycle including charging.

What is a vision-language-action model?

A vision-language-action model, or VLA, takes camera images and a natural-language instruction and outputs robot actions. Humanoids typically split it in two: a large model reasons about the task at a few hertz, and a small, fast policy converts that intent into motor commands at 100 Hz or more. Figure’s Helix, with its 7-billion-parameter reasoning layer and fast visuomotor layer, is the best-documented example.

Are humanoid robots safe to work near people?

Not yet in the way collaborative arms are. A walking robot cannot simply stop without falling, and no humanoid has a mature, widely accepted functional-safety certification. Agility’s Digit 5 is designed for field evaluation against OSHA industrial safety standards, with a dedicated safety stack, but that is an evaluation path, not a certificate. Keep humanoids in controlled zones until standards such as ISO 25785-1 mature.

2026 Humanoid Robot Benchmark: Figure 03, Optimus Gen-3, Unitree G1, Digit

2026 Humanoid Robot Benchmark: Figure 03, Optimus Gen-3, Unitree G1, Digit

Context and Background