Era0-Preview
Evaluation-only VLA model for RoboDojo simulation, contributed by RobotEra Team. Supported environment and action types: arx_x5 and ee.
Evaluate
Use the matching policy/Era0_Preview adapter from XPolicyLab. From that policy directory:
bash install.sh
export ERA0_CHECKPOINT_PATH=/path/to/era0-model
POLICY_ENV=.venv bash eval_three_seeds.sh \
--checkpoint "$ERA0_CHECKPOINT_PATH" --sim-env RoboDojo \
--execution-profile official --gpu-ids 0,1,2,3,4,5,6,7 \
--output "/path/to/eval_$(date +%F_%H%M%S)" --run
Replace RoboDojo with your simulator environment name or Python executable. Omit --run to preview. Keep the weights, model configuration, era0_adapter.json, processor/tokenizer files and norm_stats.pkl together.
Local results
The official SDK summary reports partial local Score 43.00 ± 0.68, SR 37.61 ± 0.88%, with 123/126 complete task–seed cells and 6,255/6,300 saved episodes. make_toast_random has 10/25 episodes for each seed, leaving 45 episodes missing. Incomplete cells are excluded. The current videos combine camera views and do not pass the SDK's separate-camera file inventory check.
These are local test results, not official leaderboard scores. See third-party notices for attribution.
- Downloads last month
- 161