BENCHMARK CONDITIONS STAY ATTACHED

Physical AI Benchmark Database

Compare published Physical AI benchmark results while keeping robot, environment, trial count, protocol and adaptation conditions attached to every score.

7 retained recordsLast verified: August 13, 2026Primary-source fields preserved per record

Adaptation

7 of 7 records shownFilters are local and do not create crawlable URLs

TurboVLA 216.1M

Research result

Benchmark: LIBERO (40 tasks across four suites)
Task: General manipulation suites
Model: TurboVLA 216.1M
Robot: LIBERO simulated robot
Environment: Simulation
Trials: 2,000 total (50 rollouts × 40 tasks)
Metric: Average success rate
Score: 97.7%
Adaptation: Benchmark-trained configuration
Protocol: Author evaluation; same paper protocol for listed baselines
Source: TurboVLA project

TurboVLA

Research result

Benchmark: AgileX Piper real-robot evaluation
Task: Grab the roller
Model: TurboVLA
Robot: AgileX Piper
Environment: Real lab tabletop
Trials: 40
Metric: Success rate
Score: 92.5%
Adaptation: Fine-tuned
Protocol: 65 teleoperated demonstrations for task; model then evaluated over 40 trials
Source: TurboVLA project

TurboVLA

Research result

Benchmark: AgileX Piper real-robot evaluation
Task: Move the playing card
Model: TurboVLA
Robot: AgileX Piper
Environment: Real lab tabletop
Trials: 40
Metric: Success rate
Score: 80%
Adaptation: Fine-tuned
Protocol: 65 teleoperated demonstrations for task; model then evaluated over 40 trials
Source: TurboVLA project

TurboVLA

Research result

Benchmark: AgileX Piper real-robot evaluation
Task: Press the stapler
Model: TurboVLA
Robot: AgileX Piper
Environment: Real lab tabletop
Trials: 40
Metric: Success rate
Score: 90%
Adaptation: Fine-tuned
Protocol: 65 teleoperated demonstrations for task; model then evaluated over 40 trials
Source: TurboVLA project

TurboVLA

Research result

Benchmark: AgileX Piper real-robot evaluation
Task: Stack three bowls
Model: TurboVLA
Robot: AgileX Piper
Environment: Real lab tabletop
Trials: 40
Metric: Success rate
Score: 87.5%
Adaptation: Fine-tuned
Protocol: 65 teleoperated demonstrations for task; model then evaluated over 40 trials
Source: TurboVLA project

MotuBrain

Research result

Benchmark: RoboTwin 2.0 Clean
Task: Multi-task robot control
Model: MotuBrain
Robot: RoboTwin 2.0 embodiments
Environment: Simulation benchmark
Trials: See paper protocol
Metric: Average success rate
Score: 95.8%
Adaptation: Post-trained
Protocol: Author-reported benchmark; compare only under the same RoboTwin 2.0 setup
Source: MotuBrain paper

WLA-0

Research result

Benchmark: RoboTwin 2.0 Clean
Task: Multi-task / long-horizon learning
Model: WLA-0
Robot: RoboTwin 2.0 embodiments
Environment: Benchmark environment
Trials: See paper protocol
Metric: Success rate
Score: 92.94%
Adaptation: Model-specific training
Protocol: Author-reported WLA-0 evaluation
Source: WLA-0 paper

Benchmark	Task	Model	Robot	Environment	Trials	Metric	Score	Adaptation	Protocol	Source
LIBERO (40 tasks across four suites)	General manipulation suites	TurboVLA 216.1M	LIBERO simulated robot	Simulation	2,000 total (50 rollouts × 40 tasks)	Average success rate	97.7%	Benchmark-trained configuration	Author evaluation; same paper protocol for listed baselines	TurboVLA project
AgileX Piper real-robot evaluation	Grab the roller	TurboVLA	AgileX Piper	Real lab tabletop	40	Success rate	92.5%	Fine-tuned	65 teleoperated demonstrations for task; model then evaluated over 40 trials	TurboVLA project
AgileX Piper real-robot evaluation	Move the playing card	TurboVLA	AgileX Piper	Real lab tabletop	40	Success rate	80%	Fine-tuned	65 teleoperated demonstrations for task; model then evaluated over 40 trials	TurboVLA project
AgileX Piper real-robot evaluation	Press the stapler	TurboVLA	AgileX Piper	Real lab tabletop	40	Success rate	90%	Fine-tuned	65 teleoperated demonstrations for task; model then evaluated over 40 trials	TurboVLA project
AgileX Piper real-robot evaluation	Stack three bowls	TurboVLA	AgileX Piper	Real lab tabletop	40	Success rate	87.5%	Fine-tuned	65 teleoperated demonstrations for task; model then evaluated over 40 trials	TurboVLA project
RoboTwin 2.0 Clean	Multi-task robot control	MotuBrain	RoboTwin 2.0 embodiments	Simulation benchmark	See paper protocol	Average success rate	95.8%	Post-trained	Author-reported benchmark; compare only under the same RoboTwin 2.0 setup	MotuBrain paper
RoboTwin 2.0 Clean	Multi-task / long-horizon learning	WLA-0	RoboTwin 2.0 embodiments	Benchmark environment	See paper protocol	Success rate	92.94%	Model-specific training	Author-reported WLA-0 evaluation	WLA-0 paper

Methodology

Scores are never treated as directly comparable when robots, environments, trial counts, reset rules, benchmark versions or adaptation protocols differ. The table preserves those conditions beside each number.

Included

Published model, robot, benchmark, deployment or failure evidence that can be tied to a retained source.

Not inferred

Unknown autonomy, latency, trial counts, deployment scale and failure causes remain unknown when the source does not disclose them.

Update policy

Versions and historical records remain distinguishable. A new release does not silently overwrite an older model or evaluation condition.