Environment leaderboard · v0.3.0
PickPlace v3
Goal-observable grasp, lift, and placement with dense reward over a fixed 500-step horizon.
- Distribution
- MetaWorld
- Primary metric
- assessment_mean_return
- Direction
- Higher is better
- Record date
- 11 Aug 2026
Choose experiment setup
⌄
Selected setup
128 train / 32 assessment
train-128-assess-32-v1Strictly exhaust 128 training Episodes, select with 32 Validation Episodes per candidate, then measure the selected Program on 32 held-out Assessment Episodes.
- Training Episodes
- 128
- Max Submissions
- 8
- Validation Episodes
- 32
- Assessment Episodes
- 32
Configuration details
- run seed
- 20260806
- training split
- train
- Training Episodes
- 128
- episode pool size
- 128
- Max Submissions
- 8
- max episodes per submission
- 32
- Finish Policy
- require budget exhaustion
- validation split
- validation
- Validation Episodes
- 32
- validation max candidates
- 2
- assessment split
- test
- Assessment Episodes
- 32
- episode timeout seconds
- 120
- agent timeout seconds
- 7200
Raw Assessment score and rank
1GPT-5.6 SolxhighCodex4471.544
2GPT-5.6 TerraxhighCodex4017.798
3GPT-5.6 LunaxhighCodex3728.949
Top 3 policy rollouts
The final selected Programs on the same held-out Assessment Episode.
GPT-5.6 SolCodex · xhigh
4471.544
GPT-5.6 TerraCodex · xhigh
4017.798
GPT-5.6 LunaCodex · xhigh
3728.949