Read a RoboTwin 2.0 leaderboard row in two steps: identify its training track, then select the evaluation condition that answers your question. Co-train represents one policy trained jointly across tasks; Single represents task-specific fine-tuned checkpoints. The clean2clean and clean2random columns then show evaluation in clean and randomized environments after training on the specified clean demonstrations. The default average combines those two scores.
The official leaderboard’s Benchmark Setting fixes the training data at 50 demo_clean trajectories for each of 50 tasks—2,500 demonstrations in total—on Aloha-AgileX. It specifies 100 evaluation trials per task under each of demo_clean and demo_randomized.
One policy across tasks or a checkpoint for each task
Co-train and Single-task SFT entries share the board, but the object being evaluated differs. In the co-train track, one policy is jointly trained on all 50 tasks. A Single entry uses a checkpoint fine-tuned for each task. Reading only the method name and aggregate score hides that distinction.
Imagine selecting a policy for a system that switches among several tasks. A co-trained result is directly relevant to using a shared policy across the benchmark’s tasks. A set of single-task checkpoints is relevant to an arrangement in which the system selects the appropriate specialized checkpoint. That selection and checkpoint management become part of the deployment design.
For a model comparison, start with rows in the same track. This holds the shared-versus-specialized training arrangement more nearly constant. A cross-track comparison can still be useful, but it answers a broader question about the complete training and deployment approach. Record the track beside the result so that a reader can see which arrangement produced it.
The fixed demonstration set describes this benchmark’s training setup. For a foundation policy, the method’s technical report is the place to examine earlier pretraining and the fine-tuning procedure. Those details explain how the checkpoint arrived at the benchmark, while the Track column explains how it is used across the benchmark’s tasks.
The ranking control selects a view of two conditions
The leaderboard offers three ranking choices. They correspond to two evaluation conditions and one arithmetic summary:
| Ranking choice | What is evaluated or calculated | Question it helps answer |
|---|---|---|
clean2clean |
Clean-trained policy evaluated in demo_clean |
How does it perform in the clean evaluation condition? |
clean2random |
Clean-trained policy evaluated in demo_randomized |
How does it perform when the evaluation environment is randomized? |
Average (c2c+c2r) |
(clean2clean + clean2random) / 2 |
How does it score when the two conditions receive equal weight? |
Selecting clean2random changes the evaluation column used for sorting. It does not change the stated training data into randomized demonstrations. Likewise, Average is a calculated view of the two conditions, rather than an additional training or evaluation regime.
The page labels the clean condition “easy” and the randomized condition “hard.” Those labels identify views within this protocol. For a concrete application, the useful follow-up is which randomized conditions resemble the changes your system will encounter. The aggregate randomized score is a starting point for that investigation.
Equal averages can favor different operating conditions
Consider these invented scores on a 0–100 scale, with both entries in the same co-train track. They illustrate the board’s arithmetic, not actual model results.
| Hypothetical policy | clean2clean | clean2random | Average |
|---|---|---|---|
| A | 90 | 50 | 70 |
| B | 80 | 60 | 70 |
The average gives a tie. Sorting by the clean condition places A first; sorting by the randomized condition places B first. Each ordering answers a different question using the same underlying pair of scores.
If the intended system mainly encounters the randomized condition, B’s higher value in that column is the relevant starting point. If the clean condition is the immediate target, A’s value is higher. Equal weighting in the default average is a reporting choice; the table makes its effect visible rather than concealing the two profiles in a single rank.
Next, open the per-task results for the tasks that matter. The board says these columns follow the selected ranking view and marks task-specific SFT methods as Single. Keeping the selected condition and the track visible at that level helps distinguish a broadly shared policy from a strong collection of specialists. It also shows where a favorable overall result is concentrated.
When saving a comparison, retain the method, checkpoint or report reference, track, selected ranking view and both condition scores. Another reader can then reconstruct why you preferred a row even if the board’s default ordering changes. The robot VLA evaluation guide explains the trial and configuration records that give those summary numbers their context.
This article uses the leaderboard’s published benchmark settings and column definitions, checked October 16, 2026. No RoboTwin training or evaluation was run for this article.