Robometer: Scaling General-Purpose Robotic Reward Models via Trajectory Comparisons

Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• Robometer's dual-objective training framework combining intra-trajectory progress supervision with inter-trajectory preference supervision enables effective learning from both expert and failed trajectories.method 92%
• Robometer outperforms prior methods with an average 14% improvement in reward rank correlation and 32% improvement in distinguishing successful trajectories.result 90%
• The RBM-1M dataset contains over one million trajectories spanning diverse robot embodiments and tasks, including substantial suboptimal and failure data.technical 95%
• Traditional reward modeling approaches that predict absolute task progress from expert demonstrations scale poorly to large-scale datasets with abundant failed and suboptimal trajectories.limitation 88%
• Robometer learns more generalizable reward functions than prior methods across benchmarks and real-world evaluations.result 87%
• The system depends on a massive dataset for training, which might not be as accessible or economical for smaller applications.limitation 85%
• Robometer improves robot learning performance across a diverse set of downstream applications.result 86%
• The approach introduces global supervision through pairwise trajectory comparisons to improve generalization and scale reward learning.technical 89%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card