Explore the research

Robotics

Can robots complete real tasks in unfamiliar places, reliably and without rescue?

29%

field summary estimate

6 outcomes · Assessed November 7, 2026 · Low confidence

The field estimate averages its outcome scores. It is a reading aid, not a measured fraction of the field that AI manages.

How the estimate built up.

Each band is one outcome's share of the field estimate, so the top line is the estimate itself. Every field starts at 0% on December 31, 2022. Dates before the latest are retrospective estimates. Point at a band to see that outcome's score on each date.

Useful productivity30%
Skill acquisition30%
Reliability30%
Intervention rate30%
Generalization35%
Complete task success20%

View exact historical values and reconstruction notes

What we measure

Each percentage is a judgment about progress toward independent advancement. Open the evidence to inspect the basis for it.

Complete task success

20%estimated progress

What we track

Fully completed tasks / all assigned trials, judged against observable end states rather than partial motion success.

Why this estimate?

A structured long-horizon evaluation supports bounded task assessment, but partial credit is not complete physical-task success and no verified whole-task rate was extracted.

Advancements

Winning 2025 BEHAVIOR Challenge policy reaches 26% q-score on 50 tasksDecember 7, 2025. The first-place solution in the 2025 BEHAVIOR Challenge, a benchmark of 50 long-horizon household tasks in photo-realistic simulation, achieved a 26% q-score across all 50 tasks on both the public and private leaderboards.
Humanoid HMND 01 exceeds 90% pick-and-place success in Siemens factory trialApril 20, 2026. In a two-week trial in January 2026 at Siemens' Erlangen plant, Humanoid's wheeled HMND 01 Alpha robot destacked totes with a pick-and-place success rate above 90%, at 60 container moves per hour, over more than eight hours of autonomous operation.
Gemini Robotics 2 reports per-task success from 32% to 92% on dexterous tasksJuly 30, 2026. Google DeepMind reported multi-finger task success on an Apollo humanoid with Sharpa hands of 92% for unscrewing a bulb, 36% for screwing one in, 44% for tying a trash bag, 32% for a dustpan task and 40% for a ziplock bag; whole-body pick-up averaged 68.4% from a table, 45.7% from the floor and 76.3% from a shelf.
BEHAVIOR 2026 challenge tests long household tasks2026 challenge; accessed November 7, 2026. BEHAVIOR 2026 evaluates long-horizon household activities using average partial-credit BDDL task success.

Read the dated findings and sources

Status: Benchmark protocol; whole-task rate not verified. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

2026 challenge; accessed November 7, 2026: BEHAVIOR 2026 evaluates long-horizon household activities using average partial-credit BDDL task success. BEHAVIOR 2026 official challenge.

Interpretation and tracking value

Its primary aggregate is not the fraction of trials completing every required end state. A current whole-task completion rate was not verified from the accessible challenge page.

Next comparable measurement

Archive trial-level results and count only complete end states; report successes/all assignments separately from partial credit and simulation versus physical tests. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Generalization

35%estimated progress

What we track

Success on unseen buildings, object variants, layouts, and disturbances without retraining for the individual test.

Why this estimate?

Cross-robot and compositional demonstrations support useful transfer. Coaching, related training episodes and selected tasks limit claims of independent unfamiliar-environment performance.

Advancements

RT-2 vision-language-action model generalizes to novel objects in 6,000 trials28 July 2023. The RT-2 vision-language-action model, evaluated in 6,000 robot trials, showed significantly improved generalization to novel objects and could follow commands absent from the robot training data, such as placing an object onto a particular number or icon.
Open X-Embodiment: RT-X shows positive transfer across 22 robotsOctober 13, 2023. A 21-institution collaboration released a dataset from 22 different robots covering 527 skills and 160,266 tasks, and showed that an RT-X model trained on it exhibits positive transfer, improving multiple robots' capabilities using experience from other platforms.
pi0.5 performs long-horizon cleaning tasks in entirely new homes22 April 2025. The pi0.5 vision-language-action model, co-trained on data from multiple robots, web data and semantic prediction, performed long-horizon dexterous tasks such as cleaning a kitchen or bedroom in entirely new homes, which the authors describe as a first for an end-to-end learned system.
pi0.7 transfers skills across robotsApril 16, 2026. pi0.7 demonstrates compositional tasks and cross-robot transfer; novel appliance use can improve with step-by-step language coaching.

Read the dated findings and sources

Status: Developer demonstration of transfer. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

April 16, 2026: pi0.7 demonstrates compositional tasks and cross-robot transfer; novel appliance use can improve with step-by-step language coaching. Physical Intelligence: pi0.7.

Interpretation and tracking value

Coached performance is not unassisted zero-shot success. Existing related training episodes and test-time guidance must be disclosed.

Next comparable measurement

Freeze unseen sites, layouts and objects; report complete success before and after coaching, with interventions and training overlap documented. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Intervention rate

30%estimated progress

What we track

Human rescues, remote-control episodes, and instruction corrections per 100 tasks, with labor minutes.

Why this estimate?

A laboratory executes complex workflows, yet people still replenish supplies and maintain equipment. Missing complete intervention logs prevent high autonomy credit.

Advancements

Aurora begins commercial driverless trucking between Dallas and HoustonMay 1, 2025. Aurora started commercial freight hauls with no driver on board between Dallas and Houston for Uber Freight and Hirschbach Motor Lines, completing more than 1,200 driverless miles at launch after about 3 million supervised autonomous miles and more than 10,000 customer loads.
Tesla starts Austin robotaxi rides with no safety monitor in the carJanuary 22, 2026. Tesla began public robotaxi rides in Austin without a human safety monitor inside the vehicle; its autonomy lead said the rollout started with a few unsupervised vehicles mixed into the broader fleet, while the Bay Area service kept safety drivers.
AutoDNA lab integrates more than 20 instrument typesNovember 4, 2026. AutoDNA integrates more than 20 instrument types; people still perform maintenance and replenish reagents and labware.

Read the dated findings and sources

Status: Autonomous laboratory workflow; support-rate denominator unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

November 4, 2026: AutoDNA integrates more than 20 instrument types; people still perform maintenance and replenish reagents and labware. AutoDNA, Nature.

Interpretation and tracking value

The report supports substantial automation but does not establish rescues or human minutes per 100 tasks for this project's metric.

Next comparable measurement

Log routine replenishment, setup, troubleshooting and remote control separately; report interventions and labor per 100 complete protocols. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Reliability

30%estimated progress

What we track

Operating hours between failures, recovery success, damage events, and task abandonment rate.

Why this estimate?

A large number of executed steps shows operational capability, but does not measure attempted steps, failures or hours between failures.

Advancements

Swiss Re study: Waymo has 88% fewer property damage claims over 25.3M milesDecember 19, 2024. A Swiss Re study of 25.3 million fully autonomous Waymo miles found 88% fewer property damage claims and 92% fewer bodily injury claims than the overall driving population: 9 property damage and 2 bodily injury claims against an expected 78 and 26 for human drivers.
pi*0.6 runs espresso making from 5:30am to 11:30pm with halved failure rateNovember 17, 2025. Physical Intelligence reported that its pi*0.6 model, trained with on-robot reinforcement learning (RECAP), made espresso drinks from 5:30am to 11:30pm, more than doubled throughput and cut failure rates by 2x or more on some of the hardest tasks, with each challenging task stage above 90% success.
Waymo reports 95% fewer serious-injury crashes over 271.3M driverless milesData through June 2026. Waymo's Safety Impact hub reports 271.3 million rider-only miles through June 2026 with 95% fewer serious-injury-or-worse crashes (55 fewer), 82% fewer injury-causing crashes (841 fewer) and 93% fewer pedestrian injury crashes (87 fewer) than human benchmarks in the same operating areas.
IIHS: Waymo driverless cars have 68% lower crash rate than human driversJuly 23, 2026. An Insurance Institute for Highway Safety study of about 50 million Waymo driverless miles in San Francisco, Phoenix, Los Angeles and Austin found a 68% lower police-reportable crash rate per mile than human drivers, with 81% fewer injury crashes; the rate was 35% lower in San Francisco and 4% higher in Austin, where the sample was small.
AutoDNA executes more than 11,000 experimental stepsNovember 4, 2026. AutoDNA reports executing more than 11,000 experimental steps.

Read the dated findings and sources

Status: Large execution demonstration; failure-rate metric unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

November 4, 2026: AutoDNA reports executing more than 11,000 experimental steps. AutoDNA, Nature.

Interpretation and tracking value

Steps executed are not operating hours between failures or a denominator of attempted steps. A successful demonstration cannot establish a zero-failure rate.

Next comparable measurement

Publish operating hours, attempted/completed steps, recoveries, damage and abandonment; include maintenance downtime and confidence intervals. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Skill acquisition

30%estimated progress

What we track

Demonstrations, robot-hours, and human hours required to learn a new task to a prespecified reliability target.

Why this estimate?

Learning from language coaching supports new-task acquisition with less teleoperation. Human teaching effort and time to a fixed reliability level remain unknown.

Advancements

RoboCat adapts to new tasks and robots with 100 to 1,000 examplesJune 20, 2023. The RoboCat generalist manipulation agent adapted to new tasks and robots zero-shot or with only 100 to 1,000 examples for the target task, evaluated in simulation and on three real robot embodiments, and could generate its own data for later training iterations.
Mobile ALOHA learns household mobile manipulation from 50 demonstrations per taskJanuary 4, 2024. The Mobile ALOHA system learned tasks such as sauteing and serving shrimp, storing heavy pots in a cabinet, and calling and entering an elevator from 50 demonstrations per task; the abstract states that co-training with existing static ALOHA data can increase success rates by up to 90%.
HIL-SERL reaches near-perfect success with 1 to 2.5 hours of real-robot trainingOctober 29, 2024. A human-in-the-loop reinforcement learning system learned dexterous tasks including dynamic manipulation, precision assembly and dual-arm coordination to near-perfect success rates within 1 to 2.5 hours of training, with an average 2x higher success rate and 1.8x faster execution than imitation-learning baselines and prior RL approaches.
pi0.7 learns an appliance task from language coachingApril 16, 2026. The developer reports training a high-level policy from repeated language coaching to execute an appliance task without additional teleoperation.
Gemini Robotics On-Device 2 adapts to new bi-arm robots with under 200 examplesJuly 30, 2026. Google DeepMind stated that its Gemini Robotics On-Device 2 model adapts to new bi-arm embodiments in a few hours, typically with fewer than 200 examples.

Read the dated findings and sources

Status: Learning from language coaching demonstrated. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

April 16, 2026: The developer reports training a high-level policy from repeated language coaching to execute an appliance task without additional teleoperation. Physical Intelligence: pi0.7.

Interpretation and tracking value

No complete demonstrations/robot-hours/human-hours-to-fixed-reliability baseline is established by that example. No new teleoperation does not mean no human teaching.

Next comparable measurement

Track every demonstration and coaching minute; evaluate held-out trials until a prespecified success rate is reached, counting unsuccessful training runs. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Useful productivity

30%estimated progress

What we track

Accepted physical outputs per hour and fully loaded dollars per output, including setup, maintenance, supervision, and depreciation.

Why this estimate?

Automated characterization increases specialized throughput, but accepted functional output per full dollar and unsupervised operation across settings are unmeasured.

Advancements

Amazon deploys its one millionth warehouse robotJune 30, 2025. Amazon announced its one millionth robot, delivered to a fulfillment center in Japan, across a network of more than 300 facilities, and introduced the DeepFleet foundation model, which it expects to improve robot fleet travel time by 10%.
Figure 02 loads 90,000 parts over 1,250 hours at BMW SpartanburgNovember 19, 2025. Figure reported that Figure 02 humanoids ran 10-hour weekday shifts during an 11-month deployment at BMW Group Plant Spartanburg, loading more than 90,000 sheet-metal parts over more than 1,250 hours of runtime and contributing to more than 30,000 X3 vehicles; the forearm was the top hardware failure point.
Agility's Digit moves over 100,000 totes at GXO facilityNovember 20, 2025. Agility Robotics reported that its Digit humanoid robot had moved more than 100,000 totes in commercial deployment at GXO's Flowery Branch facility.
Waymo reaches 500,000 paid robotaxi rides per week in 10 US citiesMarch 27, 2026. Waymo stated it was providing 500,000 paid robotaxi rides per week across 10 US cities, a tenfold increase from 50,000 per week in May 2024, with a fleet of over 3,000 vehicles.
SAPP automates testing of hundreds of protein designs a dayAugust 20, 2026. SAPP reports hundreds of protein designs characterized per day in an automated experimental workflow.

Read the dated findings and sources

Status: Physical workflow throughput proxy. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

August 20, 2026: SAPP reports hundreds of protein designs characterized per day in an automated experimental workflow. Protein design validation platform, Nature Communications.

Interpretation and tracking value

This is a specialized laboratory throughput result, not general household/manufacturing productivity or fully loaded dollars per accepted functional output.

Next comparable measurement

Measure externally accepted outputs per operating hour and total cost including supervision, failures, maintenance, consumables and depreciation. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note