Event2Reward: Learning Robotic Rewards from Progress Relations
Event2Reward: Learning Robotic Rewards from Progress Relations
Event2Reward is a reinforcement learning method that constructs dense shaping rewards for robots by learning from progress relations in demonstrations. It uses normalized temporal labels and within-trajectory progress relations to improve policy success across real-world tasks, outperforming sparse-reward baselines by significant margins.
This paper introduces Event2Reward, a method for learning robotic rewards by leveraging progress relations from demonstration data. It addresses the limitations of using absolute progress labels by incorporating both absolute and relative progress supervision, leading to improved policy success in real-world tasks.
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
number of tasks in LIBERO suites
40
View original evidence
ing framewise absolute progress labels. Under the same training budgets, Event2Reward improves average policy success by 18.16 percentage points over a baseline using only sparse task rewards across 40 tasks in four LIBERO suites. On real robots, average success after 30 minutes of wall-clock training reaches 60.0% on three precise insertion tasks and 90.0% on three object manipulation tasks, exceeding the sparse-rew
Open source S5number of raw real-robot execution records
≈100,000
View original evidence
d, which uses normalized temporal labels to anchor the progress scale and these within-trajectory relations to constrain local changes, combining absolute and relative progress supervision. Its progress predictions serve as potentials for reward shaping. We curate an event-annotated dataset from approximately 100K raw real-robot execution records, covering successful, suboptimal, and failed executions, without requir
Open source S4
- paper: not_reported
- code: not_reported
- weights: not_reported
- dataset: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
ing framewise absolute progress labels. Under the same training budgets, Event2Reward improves average policy success by 18.16 percentage points over a baseline using only sparse task rewards across 40 tasks in four LIBERO suites. On real robots, average success after 30 minutes of wall-clock training reaches 60.0% on three precise insertion tasks and 90.0% on three object manipulation tasks, exceeding the sparse-rew
Open source S5
ard baseline under the same budget by 36.7 and 23.3 percentage points, respectively. Ablation studies further show that absolute progress supervision and relative progress relations are complementary, with joint supervision yielding better downstream policy learning.
Open source S6
d, which uses normalized temporal labels to anchor the progress scale and these within-trajectory relations to constrain local changes, combining absolute and relative progress supervision. Its progress predictions serve as potentials for reward shaping. We curate an event-annotated dataset from approximately 100K raw real-robot execution records, covering successful, suboptimal, and failed executions, without requir
Open source S4
Source：arXiv Robotics — research abstracts · arxiv.org