Experience can change.
The simulator and Python state persist across submissions. The agent revises control code, Skills, and Lessons, and can reset the same scene to try again.
Method & Results
Learn in Playground.
Transfer to Test.
Explore through code, retain Skills & Lessons, then freeze that experience for fresh agents in independent scenes.
Explore through code.
The coding agent writes and executes Python in a persistent robot session. It observes what happened, revises the next segment, and can reset in the Playground.
Programs are
the action interface.
The agent emits executable Python segments. A segment can combine numerical reasoning, conditionals, loops, and reusable routines before calling the robot’s native controller.
RoboDojo exposes RGB images and robot state. The native controller handles inverse kinematics and interpolation; object ground-truth poses, depth, and camera calibration are unavailable.
THE FIXED ENVIRONMENT INTERFACE4 PRIMITIVES
get_instruction()Read the taskget_observation()Inspect the current statestep(action)Execute native controlreset()Playground onlyWhere code fits in robot control.
Knowledge is frozen.
Each scene gets a fresh agent, workspace, Python namespace, and simulator. Local programs still adapt through feedback. Inherited files are read-only, reset is disabled, and Test scenes share no new experience.
The model and interface.
Model weights, task goals, native primitives, and success conditions stay fixed. Each Playground chain and Test episode allows up to 100 code executions, subject to the task’s native-action allowance.
Key Findings
01 Experience improves execution under both inputs.
With task descriptions alone or with the wiki, frozen Skills & Lessons raise success by more than 21 percentage points on matched scenes.
8.50 → 30.17%Success rate · Astra/xhigh
Task description only
Experience improves execution under both inputs.
With task descriptions alone or with the wiki, frozen Skills & Lessons raise success by more than 21 percentage points on matched scenes.
Task description only
RoboDojo
SUCCESS RATE (%)0 → 100
INHERITED EXPERIENCE+21.67pp
success rate
95% CI: 12.33 to 31.39 pp| Experience | Score ↑ | Success rate ↑ |
|---|---|---|
| No experience | 14.77 | 8.50% |
| Skills & Lessons | 38.64 | 30.17% |
SUCCESS RATE (%)0 → 100
INHERITED EXPERIENCE+21.72pp
success rate
| Experience | Score ↑ | Success rate ↑ |
|---|---|---|
| No experience | 20.67 | 13.36% |
| Skills & Lessons | 42.42 | 35.08% |
02 Learn how to act—and when to act.
A reusable grasp enables repeated placements; an observation lesson helps the agent wait for a complete sequence. Precise contact still limits task completion.
0 → 100%Stack blocks · paired success rate
Five Test scenes
Learn how to act—and when to act.
A reusable grasp enables repeated placements; an observation lesson helps the agent wait for a complete sequence. Precise contact still limits task completion.
Five Test scenes
Both tasks become successful with inherited control procedures.
The agent observes the new sequence before reproducing it.
Contact still matters.
Tube and key insertion gain partial Score but remain at 0% success. In the key case, the handover fails before insertion even begins.
PROCEDURES + CONDITIONS
The tic-tac-toe agent reuses rim_pick to retain a ring and complete the board. The sorting agent waits to observe an order different from Playground. These traces show how Skills and Lessons are used; they do not isolate either artifact’s causal effect.
03 Experience transfers across models and tasks.
Weaker executors use knowledge they struggle to acquire. On twelve target tasks, experience from other tasks also helps without new target-task exploration.
+8.86 / +5.50 ppSol / Luna success-rate gains
Compared with no experience
Experience transfers across models and tasks.
Weaker executors use knowledge they struggle to acquire. On twelve target tasks, experience from other tasks also helps without new target-task exploration.
Compared with no experience
No experienceOwn harnessAstra/xhigh harness
Astra
xhighGPT-6 Astra
SUCCESS RATE0 → 40%
Astra
mediumGPT-6 Astra
SUCCESS RATE0 → 40%
Sol
xhighGPT-5.6 Sol
SUCCESS RATE0 → 40%
Luna
xhighGPT-5.6 Luna
SUCCESS RATE0 → 40%
PRODUCER QUALITY MATTERS
Sol and Luna gain no task-completion benefit from their own harnesses. Astra-produced experience raises their success rates by 8.86 and 5.50 percentage points over no experience. Astra/medium also benefits from the higher-effort producer.
ACROSS TASKSUseful operations travel
Useful operations travel
beyond task labels.
On twelve fixed targets, related sources add 14.72 pp success and mismatched sources add 16.67 pp, without target-task exploration. The ordering reverses across individual tasks: applicability depends on the retained operations, not just label similarity.
| Experience source | Score | Success |
|---|---|---|
| No experience | 7.35 | 1.67% |
| Same-task experience | 31.78 | 23.06% |
| Related-task source | 21.53 | 16.39% |
| Mismatched source | 26.00 | 18.33% |
04 Unfinished exploration can still leave useful knowledge.
Image-conditioned conveyor pickup never succeeds in Playground, yet the retained operations help fresh agents complete the task on new scenes.
0 → 40%Conveyor pickup · Test success
Astra/xhigh · five new scenes
Unfinished exploration can still leave useful knowledge.
Image-conditioned conveyor pickup never succeeds in Playground, yet the retained operations help fresh agents complete the task on new scenes.
Astra/xhigh · five new scenes
FROM PARTIAL PROGRESSLearn an operation.
Learn an operation.
Reuse it to finish.
The Playground task remains unfinished after 100 executions. Execution 74 still establishes a tracked grasp and test lift in 31 native actions. Frozen Skills & Lessons carry that local progress into Test.
| Test experience | Score | Success |
|---|---|---|
| No experience | 0 | 0% |
| Skills & Lessons | 40 | 40% |
Capture and carry.
Moving-target capture and contact-guarded transport provide reusable parts of the task.
Track the moving target.
Lessons tell the next agent to update its targets as the conveyor moves. Partial success in an operation can help even without a complete successful episode.
05 Structured experience costs less to reuse.
Skills & Lessons achieve similar observed success to raw history at lower Test expense, while costing more than execution without experience.
53.0% lowerEstimated Test expense vs. raw history
Acquisition excluded
Structured experience costs less to reuse.
Skills & Lessons achieve similar observed success to raw history at lower Test expense, while costing more than execution without experience.
Acquisition excluded
8.50% success · 14.77 Score
29.78% success · 36.41 Score
30.17% success · 38.64 Score
COMPACT EXPERIENCE
Skills & Lessons use 11.3% fewer total Test tokens than raw history (3.958M vs. 4.460M per episode), with similar observed success: 30.17% vs. 29.78%. The 87.9% reduction in uncached input explains the larger cost saving under the paper’s fixed token rates.
Full RoboDojo evaluation & leaderboard
44.48 / 37.82%Average Score / success rate
Full RoboDojo evaluation & leaderboard
Strongest on Memory and Open.
44.48Average Score
37.82%Success rate
EmbodiedEvo has the highest average Score and success rate among the paper’s seven compared systems, ahead of VPP2 by 5.22 Score points and 5.56 success-rate points. Memory and Open lead their categories; Precision and Long-Horizon remain below the strongest listed systems.
SUCCESS RATE BY CATEGORY0 → 100%
Generalization12 tasks29.67%
Precision8 tasks13.25%
Long-Horizon8 tasks23.50%
Memory6 tasks69.67%
Open8 tasks53.00%
RoboDojo leaderboard
Snapshot · Oct 9, 2026Official leaderboard
| Model / contributor | Type | Score/SR | Score/SR | Score/SR | Score/SR | Score/SR | Score/SR |
|---|---|---|---|---|---|---|---|
 EmbodiedEvoOurs Navers Lab · Einsia AIAgent | Agent | 44.48/37.82% | 34.12/29.67% | 23.65/13.25% | 37.20/23.50% | 71.90/69.67% | 55.52/53.00% |
 VPP2 RobotEraWAM | WAM | 39.26/32.26% | 36.38/29.50% | 43.03/35.17% | 56.28/37.00% | 56.01/55.56% | 4.61/4.08% |
 PhysicalRSI HKU MMLab & KAIAgent + VLA | Agent + VLA | 36.27/31.38% | 21.30/15.62% | 38.05/32.50% | 46.37/37.33% | 46.74/46.56% | 28.88/24.92% |
 Awomo-0.5 AwomoWAM | WAM | 35.34/29.64% | 30.31/23.84% | 39.65/33.50% | 40.36/26.75% | 42.27/41.22% | 24.09/22.92% |
 Simate-beta SimateVLA | VLA | 33.95/27.96% | 35.09/27.95% | 34.35/26.92% | 57.84/43.42% | 33.33/33.00% | 9.12/8.50% |
 InternW0-Δ Shanghai AI LabWAM | WAM | 30.77/23.91% | 30.09/22.78% | 31.98/23.25% | 46.29/29.33% | 34.67/34.00% | 10.84/10.17% |
 GPT-6-Astra RoboDojo TeamAgent | Agent | 28.97/22.48% | 33.36/30.50% | 12.65/4.00% | 21.45/8.25% | 43.04/38.67% | 34.36/31.00% |
↑ Higher is better · Score / success rate (%)Scroll to compare all capabilities
Explore the evidence.
The study evaluates one generation of experience acquisition and inheritance in RoboDojo simulation. Sustained improvement across generations and real-hardware performance remain open questions.
AT A GLANCE
- Environment
- RoboDojo simulation
- Main comparison
- 42 tasks · 210 episodes per condition
- Full evaluation
- 2,100 episodes · one seed
- Transferred knowledge
- Frozen Skills & Lessons
- Updated through experience
- External control knowledge
- Held fixed
- Model weights & robot interface