Game-Enhanced Embodied Learning with Event and Foresight Pretraining
Game-Enhanced Embodied Learning with Event and Foresight Pretraining
This research paper introduces a method for enhancing embodied learning using human gameplay data. The authors study event-and-foresight pretraining for the InternVLA-A1.5 model, combining keyboard/mouse event prediction with visual supervision from a frozen video model. Using a corpus of 394 recordings across 20 games totaling 223.36 hours and 2,072,084 training anchors, the study demonstrates that after 12,400 pretraining updates, the model achieves an aggregate success rate increase from 85.90% to 90.55% across four LIBERO suites. The results highlight the practical value of game-pretraining while noting heterogeneous task outcomes and unisolated effects of optimization and training variability.
The study reports a 4.65 percentage point increase in aggregate success rates across four LIBERO suites for the InternVLA-A1.5 model after game-enhanced pretraining, though the authors note that optimization effects and training variability were not isolated.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
training anchors
2,072,084
View original evidence
s, and an attention mask prevents teacher-forced future actions from entering foresight queries. The selected corpus contains 394 recordings across 20 games, with 223.36 hours of recorded timestamp span and 2,072,084 training anchors. After 12,400 game-pretraining updates, we compare the resulting initialization with the public base under the same suite-specific robot fine-tuning protocol. Across four LIBERO suites,
Open source S3recorded timestamp span
223.36 hours
View original evidence
s, and an attention mask prevents teacher-forced future actions from entering foresight queries. The selected corpus contains 394 recordings across 20 games, with 223.36 hours of recorded timestamp span and 2,072,084 training anchors. After 12,400 game-pretraining updates, we compare the resulting initialization with the public base under the same suite-specific robot fine-tuning protocol. Across four LIBERO suites,
Open source S3game-pretraining updates
12,400
View original evidence
s, and an attention mask prevents teacher-forced future actions from entering foresight queries. The selected corpus contains 394 recordings across 20 games, with 223.36 hours of recorded timestamp span and 2,072,084 training anchors. After 12,400 game-pretraining updates, we compare the resulting initialization with the public base under the same suite-specific robot fine-tuning protocol. Across four LIBERO suites,
Open source S3success count (Spatial)
1,718
Reported trials
View original evidence
aggregate success increases from 1,718/2,000 (85.90%) to 1,811/2,000 (90.55%), a gain of 4.65 percentage points. Suite gains are 9.2, 6.8, 0.2, and 2.4 points on Spatial, Object, Goal, and Long. The 40-task breakdown shows 24 improvements, 10 declines, and six ties. These results support the practical value of the complete game-pretraining stage in this evaluated setting, while revealing heterogeneous task outcomes.
Open source S4success count (Object)
1,811
Reported trials
View original evidence
aggregate success increases from 1,718/2,000 (85.90%) to 1,811/2,000 (90.55%), a gain of 4.65 percentage points. Suite gains are 9.2, 6.8, 0.2, and 2.4 points on Spatial, Object, Goal, and Long. The 40-task breakdown shows 24 improvements, 10 declines, and six ties. These results support the practical value of the complete game-pretraining stage in this evaluated setting, while revealing heterogeneous task outcomes.
Open source S4
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
s, and an attention mask prevents teacher-forced future actions from entering foresight queries. The selected corpus contains 394 recordings across 20 games, with 223.36 hours of recorded timestamp span and 2,072,084 training anchors. After 12,400 game-pretraining updates, we compare the resulting initialization with the public base under the same suite-specific robot fine-tuning protocol. Across four LIBERO suites,
Open source S3
aggregate success increases from 1,718/2,000 (85.90%) to 1,811/2,000 (90.55%), a gain of 4.65 percentage points. Suite gains are 9.2, 6.8, 0.2, and 2.4 points on Spatial, Object, Goal, and Long. The 40-task breakdown shows 24 improvements, 10 declines, and six ties. These results support the practical value of the complete game-pretraining stage in this evaluated setting, while revealing heterogeneous task outcomes.
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org