IJRES IF 7.52Published April 2025Real Research Data← Portfolio
Live Research Results Dashboard
Real results from published research and independent projects — not mock data, not placeholder numbers.
Every metric on this page comes directly from actual experiments, training runs, and test logs.
During early training the PPO agent discovered a specification gaming exploit: spinning finger actuators in place to maximise rotation velocity reward without actually grasping the pen. This is a direct instance of reward misspecification. Mitigation: phase-based reward decomposition (Grasp → Rotate → Hold) with gated activation. Result: agent converged to genuine manipulation behaviour.
PPO Training Reward Progression
Episode reward over 73,728 timesteps. Peak 18,007 at Phase 3 convergence. Mid-training dip reflects reward hacking phase.
Touch Sensor Activation Over Time
Normalised contact intensity per finger. Index, Thumb, and Middle fingers show progressive coordination across training.
KL Divergence
Policy update stability. Decreases from 0.020 to 0.006 — safe PPO clip range maintained throughout.
Value & Policy Loss
Value loss converges from 1.8M to 0.4M. Policy gradient loss stabilises at -0.00534.
3-Seed Reproducibility
Mean ± std across 3 independent seeds. Final: 3.118 ± 0.047 rad — statistically robust.