Find the physical experiment
Start with the hardware and task description. Record the robot version, sensors, objects and starting conditions. Separate simulation results from physical trials. Keep training environments apart from the environments used for evaluation.
This is our editorial reading method. The ACT example below shows how to recover a task definition and trial count from a published paper.
Make a small evidence card
| Field | Question |
|---|---|
| Inputs | Which cameras, sensors, prompts or state estimates reach the controller? |
| Outputs | Does the model issue joint commands, target poses or higher-level actions? |
| Data | What training data and evaluation split are described? |
| Result | How many successes occurred in how many attempts? |
| Variation | Which objects, positions or environments changed? |
| Failure | What went wrong and what recovery was permitted? |
Keep the denominator
The 2023 ACT paper reports 96 percent for the final Slot Battery stage. Section V-C gives 25 physical trials from one trained seed. That implies 24 successes. [1]
| Evidence-card field | Slot Battery in ACT |
|---|---|
| Hardware and task | Two ALOHA arms place and fully insert a battery into a remote control |
| Evaluation variation | Initial object placement varies along a 15 cm line |
| Result | 96 percent final success over 25 trials for one seed |
| Scope | This task and setup, rather than every fine manipulation task |
Keep the last task stage separate from earlier progress. Grasping the battery and placing it in the slot do not establish completed insertion. Do not combine the physical count with the simulated trials in the same table. [1]
Read the definition of success before comparing two methods. Check whether both used the same objects, number of attempts, time limit and amount of human help. Leave incomparable results in separate rows.
Keep a route back to the experiment
- Save the exact title, author list and paper version.
- Keep the supplementary video and code release beside the paper.
- Record missing hardware details or evaluation settings.
- Use the conference archive to confirm the publication record.
The ACT authors provide the paper, code and demonstration videos on the ALOHA project page. Keep those materials with the evidence card. [2]
Sources and verification
- Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware ↗Tony Z. Zhao and colleagues · Read 8 October 2026
2023 v1. Table I and section V-C supply the physical Slot Battery rate and denominator.
- Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware ↗Tony Zhao, Vikash Kumar, Sergey Levine and Chelsea Finn · Read 8 October 2026
RSS 2023 project page for ALOHA and ACT. Results apply to the tested setup.
Article history
Replaced the hypothetical rate with the ACT Slot Battery table, its 25 physical trials and a derived success count.
Report a correction