robo-dopamine-rl

Overcome reward function design challenges via General Reward Model (GRM) for step-wise progress assessment. Uses multi-view observations for occlusion robustness, hop-based progress normalization, Po

Files1

View on GitHub

SKILL.md

GitHub

Overview

Addresses reward design challenges in robotic manipulation via principled progress modeling.

Core Technique

General Reward Model:

# Hop-based progress: normalized by task span
progress = model.predict_progress(observation, task_span)

# Multi-view fusion
fused_progress = aggregate_views([view1, view2, view3])

# Policy-Invariant Reward Shaping
reward = outcome_reward + policy_invariant_shaping(fused_progress)

Performance

  • 92.8% task completion accuracy
  • 95% success within 150 interactions
  • Theoretically grounded reward shaping

References

  • Step-wise progress discretization
  • Multi-view observation fusion
  • Policy-Invariant Reward Shaping

Related in other

limits-colimits

Context management for Claude Code. Hooks maintain state via ledgers and handoffs. MCP execution without context polluti...

by parcadei
3,392

ddttom-allaboutv2-claude-skills-block-collection-and-party

The Block Collection and Block Party are repositories for existing AEM blocks, build tools, code snippets, integration p...

by mattnigh
20

nodnarbnitram-claude-code-extensions-claude-skills-esphome-confi

This skill should be used when the user asks to "create an esphome config", "set up an esp32 device", "configure esphome...

by mattnigh
20