Physical AI / Data and evidence

Real-robot tests

See what Physical AI models actually demonstrated on real robots, including task, control mode, trials, failures, interventions and evidence limits.

5 documented cases
Last dataset review: 2026-08-13
Research evaluations and demos

How to read these results

Each record identifies the robot, task, test conditions, reported outcomes and source. Success rates from different tasks or experimental setups are not a common ranking. Where authors did not disclose trial counts or operator assistance, the entry retains that limitation.

Research records

Published real-robot evidence

5 of 5 records

01HUST / Huawei

TurboVLA

AgileX Piper

Four tabletop tasks roller, playing card, stapler, bowl stacking

Original evidence Research result
Environment

Lab tabletop

Trials reported

160 total

Control and autonomy

Autonomous policy execution after teleoperated training data collection

Task policy

Reported success

87.5% aggregate across the four 40-trial task sets

Human interventions

Runtime intervention count not reported

Test conditions, failures and limits
Setup
Two RGB-D camera views; 65 teleoperated demonstrations per task
Reported failures
20 total inferred exactly from published trial counts and success rates
Recovery
Not reported in aggregate
Run duration
Not reported
Study limitations
Controlled tabletop tasks on one arm; no long-duration deployment evidence
Source verified
2026-08-13
02Beijing Humanoid Robot Innovation Center

XR-1

Six robot embodiments

>120 manipulation tasks

Original evidence Research result
Environment

Real-world research evaluations

Trials reported

>14,000 rollouts reported

Control and autonomy

Learned VLA policy

Autonomous evaluation under paper protocol

Reported success

Not reduced to one cross-task figure here

Human interventions

See paper protocol

Test conditions, failures and limits
Setup
Six embodiments and heterogeneous tasks; multi-stage training
Reported failures
Model/task-specific; no single aggregate count represented here
Recovery
Not represented as a common metric
Run duration
Not disclosed as one aggregate runtime
Study limitations
Published research evaluation, not proof of unsupervised production deployment
Source verified
2026-08-13
03Tsinghua University / collaborators

RDT-1B

ALOHA dual-arm and other manipulation platforms

Bimanual manipulation, language following, unseen objects/scenes and few-shot skills

Original evidence Research result
Environment

Real robot research evaluation

Trials reported

See paper

Control and autonomy

Diffusion action policy

Autonomous policy evaluation under paper protocol

Reported success

Not collapsed into one universal number

Human interventions

Not normalized here

Test conditions, failures and limits
Setup
Fine-tuned with 6K+ ALOHA episodes after 1M+ episode multi-robot pretraining
Reported failures
See task-specific tables
Recovery
Not reported as one common metric
Run duration
Not disclosed as long-duration operation
Study limitations
Research task protocols differ from deployment workloads
Source verified
2026-08-13
04Google DeepMind

Gemini Robotics 2

Apptronik Apollo

Whole-body locomotion plus manipulation demonstrations

Original evidence Official demo
Environment

Real robot demonstrations

Trials reported

Not disclosed as one common total

Control and autonomy

VLA whole-body control

Official autonomous task demonstrations; exact supervision varies by experiment

Reported success

Task-specific official results

Human interventions

Not disclosed as one common rate

Test conditions, failures and limits
Setup
Official Google DeepMind evaluation setup
Reported failures
Not disclosed as one common total
Recovery
Demonstrated in selected scenarios; denominator not public
Run duration
Not disclosed
Study limitations
Closed model and controlled evaluation; public data does not establish fleet reliability
Source verified
2026-08-13
05Stanford / UC Berkeley / collaborators

OpenVLA 7B

Multiple manipulation robots

Language-conditioned manipulation

Original evidence Research result
Environment

Real robot research evaluations

Trials reported

Protocol-dependent

Control and autonomy

VLA policy

Autonomous policy evaluation under project protocol

Reported success

See individual project evaluations

Human interventions

Not normalized across platforms

Test conditions, failures and limits
Setup
Model pretrained on Open X-Embodiment and adapted/evaluated on target setups
Reported failures
Protocol-dependent
Recovery
Not a common reported metric
Run duration
Not long-duration deployment evidence
Study limitations
Different robots require adaptation; benchmark and real-world protocols are not interchangeable
Source verified
2026-08-13

Evidence and reporting rules

Scripted demos, teleoperation, shared autonomy, autonomous policy tests and deployments are kept separate. Training teleoperation is not mislabeled as runtime teleoperation, and no system is called fully autonomous without explicit evidence.

The records reflect their retained sources and verification dates. This page is not a live monitor of robot deployments.