How ankurjyotisarma/can-llms-help-us-in-playing-simulated-table-tennis-through-a-musculoskeletal-model works
This research codebase trains and evaluates PPO agents for musculoskeletal table tennis in MyoSuite. The principal workflow is to shape rewards around simulated play, train or fine-tune an agent, then evaluate gameplay or checkpoints and inspect reward, success-rate, and learning-curve analytics. The README documents these capabilities; excerpts show V8 environment setup and reward shaping, plus evaluation and graph-generation logic. Training-to-artifact links are documented at a high level, but their exact save/load wiring is not shown in the excerpts.
Main entry points: V8.py V8 training · 4. Detailed Evaluation using graphs Interactive analytics · Final_Model_Evaluation.py Final V8 evaluation · Learning_Curve_V8.py V8 500-game curve
Components 17
Training iterations
- Iterations V1–V6
2. Python files - V7 training
3. Fine Tuned versions V7 & V8/V7 Agent/V7.py - V8 training
3. Fine Tuned versions V7 & V8/V8 ( Iteration 8 ) Agent/V8.py - V8 training baseline
2. Python files/V8.py - Reward shaping
3. Fine Tuned versions V7 & V8/V8 ( Iteration 8 ) Agent/V8.py - PPO learner
- Model checkpoints
- Normalization stats
3. Fine Tuned versions V7 & V8/V8 ( Iteration 8 ) Agent/vec_normalize_v8.pkl
Agent evaluation
- V7 live evaluation
3. Fine Tuned versions V7 & V8/V7 Agent/Evaluate_Agent_V7_Live_Playing.py - V8 live evaluation
3. Fine Tuned versions V7 & V8/V8 ( Iteration 8 ) Agent/Evaluate_Agent_V8_Live_Playing.py - Final V8 evaluation
4. Detailed Evaluation using graphs/2. Evaluation of Final Trained Agent V8 (Iteration 8)/Final_Model_Evaluation.py
Performance analytics
- V8 500-game curve
4. Detailed Evaluation using graphs/1. Whole Learning Curve (Chcking all Check-Points)/1. MYO_AGENT_V8_LEARNING_CURVE/1. V8 Evaluation — 500 Games per Checkpoint/Learning_Curve_V8.py - V8 200-game curve
4. Detailed Evaluation using graphs/1. Whole Learning Curve (Chcking all Check-Points)/1. MYO_AGENT_V8_LEARNING_CURVE/2. V8 Evaluation — 200 Games per Checkpoint/Learning_Curve_V8.py - V7/V8 curve
4. Detailed Evaluation using graphs/1. Whole Learning Curve (Chcking all Check-Points)/2. MYO_AGENT_V7_and_V8_COMBINED_FULL_LEARNING_CURVE/MYO_CHALLENGE_AGENT_V7_and_V8_COMBINED_FULL_LEARNING_CURVE.py - Interactive analytics
4. Detailed Evaluation using graphs
Simulation environment
- MyoSuite table tennis
Other components
- Researcher
Connections 19
- Researcher → V8 training: starts training
README.md - Researcher → V8 live evaluation: evaluates agent
README.md - V8 training → Reward shaping: applies shaping
3. Fine Tuned versions V7 & V8/V8 ( Iteration 8 ) Agent/V8.py - V8 training → MyoSuite table tennis: creates environment
3. Fine Tuned versions V7 & V8/V8 ( Iteration 8 ) Agent/V8.py - V8 training → PPO learner: configures PPO
README.md - V7 training → MyoSuite table tennis: trains in
- V8 training baseline → MyoSuite table tennis: creates environment
2. Python files/V8.py - V8 training → Model checkpoints: saves checkpoints
- V8 training → Normalization stats: stores statistics
- V7 live evaluation → MyoSuite table tennis: evaluates episodes
3. Fine Tuned versions V7 & V8/V7 Agent/Evaluate_Agent_V7_Live_Playing.py - V8 live evaluation → MyoSuite table tennis: evaluates episodes
3. Fine Tuned versions V7 & V8/V8 ( Iteration 8 ) Agent/Evaluate_Agent_V8_Live_Playing.py - Final V8 evaluation → MyoSuite table tennis: runs evaluation
4. Detailed Evaluation using graphs/2. Evaluation of Final Trained Agent V8 (Iteration 8)/Final_Model_Evaluation.py - V8 500-game curve → MyoSuite table tennis: evaluates checkpoints
4. Detailed Evaluation using graphs/1. Whole Learning Curve (Chcking all Check-Points)/1. MYO_AGENT_V8_LEARNING_CURVE/1. V8 Evaluation — 500 Games per Checkpoint/Learning_Curve_V8.py - V8 200-game curve → MyoSuite table tennis: evaluates checkpoints
4. Detailed Evaluation using graphs/1. Whole Learning Curve (Chcking all Check-Points)/1. MYO_AGENT_V8_LEARNING_CURVE/2. V8 Evaluation — 200 Games per Checkpoint/Learning_Curve_V8.py - V7/V8 curve → MyoSuite table tennis: evaluates checkpoints
4. Detailed Evaluation using graphs/1. Whole Learning Curve (Chcking all Check-Points)/2. MYO_AGENT_V7_and_V8_COMBINED_FULL_LEARNING_CURVE/MYO_CHALLENGE_AGENT_V7_and_V8_COMBINED_FULL_LEARNING_CURVE.py - V8 500-game curve → Interactive analytics: generates graphs
4. Detailed Evaluation using graphs/1. Whole Learning Curve (Chcking all Check-Points)/1. MYO_AGENT_V8_LEARNING_CURVE/1. V8 Evaluation — 500 Games per Checkpoint/Learning_Curve_V8.py - V8 200-game curve → Interactive analytics: generates graphs
4. Detailed Evaluation using graphs/1. Whole Learning Curve (Chcking all Check-Points)/1. MYO_AGENT_V8_LEARNING_CURVE/2. V8 Evaluation — 200 Games per Checkpoint/Learning_Curve_V8.py - V7/V8 curve → Interactive analytics: generates graphs
4. Detailed Evaluation using graphs/1. Whole Learning Curve (Chcking all Check-Points)/2. MYO_AGENT_V7_and_V8_COMBINED_FULL_LEARNING_CURVE/MYO_CHALLENGE_AGENT_V7_and_V8_COMBINED_FULL_LEARNING_CURVE.py - Final V8 evaluation → Interactive analytics: generates graphs
4. Detailed Evaluation using graphs/2. Evaluation of Final Trained Agent V8 (Iteration 8)/Final_Model_Evaluation.py