SILGym Trainer Execution Guide

This guide provides comprehensive documentation for all possible trainer execution options and configurations for the SILGym framework.

Table of Contents

1. Basic Command Structure
2. Command-Line Arguments
3. Algorithm Configurations
4. Environment and Scenario Options
5. Lifelong Learning Algorithms
6. Complete Examples

Basic Command Structure

The basic command to run the trainer is:

python exp/trainer.py [OPTIONS]

Command-Line Arguments

Debug and Experiment Settings

Argument	Short	Type	Default	Description
--debug	-d	flag	False	Enable debug mode (sets phase_epochs=1)
--exp_id	-id	string	''	Experiment ID for tracking and organization
--seed	-seed	int	0	Random seed for reproducibility

Environment and Scenario Configuration

Argument	Short	Type	Default	Options	Description
--env	-e	string	'kitchen'	kitchen, mmworld	Environment to use
--scenario_type	-sc	string	'objective'	See Scenario Options	Type of scenario
--sync_type	-st	string	'sync'	See Sync Options	Synchronization type

Algorithm Configuration

Argument	Short	Type	Default	Options	Description
--algorithm	-al	string	'ptgm'	ptgm, buds, iscil, imanip, assil, lazysi, silc	Skill incremental learning algorithm
--lifelong	-ll	string	'append'	See Lifelong Options	Decoder lifelong learning strategy

Training Configuration

Argument	Short	Type	Default	Description
--epoch	-epoch	int	None	Number of training epochs per phase
--dist_type	-dt	string	'maha'	Distance metric type: maha, euclidean, cossim

Evaluation Settings

Argument	Type	Default	Description
--do_eval	flag	True	Perform remote evaluation after each phase
--no_eval	flag	False	Skip remote evaluation

Algorithm Configurations

1. PTGM (Prototype-based Task Generation and Matching)

• Class: PTGMConfig
• Special lifelong options:
  • notsne prefix: Disable t-SNE
  • s<num>: Set cluster number
  • g<num>: Set groups number
• Example: --algorithm ptgm --lifelong s20g20append4

2. BUDS (Behavior Unsupervised Discovery of Skills)

• Class: BUDSConfig
• Lifelong options: Standard (ft, er, append)
• Example: --algorithm buds --lifelong append4

3. IsCiL (Incremental Skill Continual Learning)

• Class: IsCiLConfig
• Special lifelong options:
  • bases<num>: Set number of prototype bases (default: 50)
• Example: --algorithm iscil --lifelong bases100

4. Imanip (Instruction Manipulation)

• Class: ImanipConfig
• Special lifelong options:
  • tr<num>: Temporal replay with buffer ratio
• Example: --algorithm imanip --lifelong tr20

5. AsSIL (Adaptive Skill Sequential Incremental Learning)

• Class: AsSILConfig
• Special lifelong format: [zero]s<decoder>p<policy>b<bases>[r<ranks>][g<groups>]
  • zero prefix: Use zero-shot agent
  • s<num>: Decoder skill count
  • p<num>: Policy skill count
  • b<num>: Number of prototype bases
  • r<num>: Optional rank dimension
  • g<num>: Optional groups/bin count
• Example: --algorithm assil --lifelong s20p10b3r4g20

6. LazySI (Lazy Skill Interface)

• Class: LazySIConfig
• Complex lifelong format: [algo_mode]/decoder_part/dec_conf/policy_algo/pol_conf
  • algo_mode: few<shot>[frac<num>] or conf<num>[_chi2|_percentile]
  • decoder_part: ptgm, buds, or semantic
  • policy_algo: ptgm, buds, instance, or static
• Example: --algorithm lazysi --lifelong conf99/ptgm/s20g20b4/ptgm/s20g20b4

7. SILC (Skill Incremental Learning with Clustering)

• Class: SILCConfig
• Same options as LazySI (refactored version)
• Example: --algorithm silc --lifelong conf99_chi2/ptgm/s20b4/instance/g20b1

Environment and Scenario Options

Kitchen Environment

Scenario Types

• kitchenem: Object-based clustering and synchronization
• kitchenex: Explicit skill incremental learning
• objective_p1, objective_p2, objective_p3: Object-based with permutations
• debug: Debug scenario (4 tasks only)

Sync Types

• sync: Standard synchronization (only option)

MMWorld Environment

Scenario Types

• mmworldem: Easy mode explicit skills

Sync Types

• sync: Standard synchronization (only option)

Lifelong Learning Algorithms

Basic Options (Available for all algorithms)

1. Fine-tuning (ft)

   • No replay or model expansion
   • Example: --lifelong ft

2. Experience Replay (er)

   • Format: er<percentage>
   • Default: 10% if no percentage specified
   • Example: --lifelong er20 (20% buffer)

3. Model Expansion (append)

   • Format: append<lora_dim>
   • Default LoRA dimension: 4
   • Example: --lifelong append16

Algorithm-Specific Options

See individual algorithm descriptions above for special lifelong learning options.

Complete Examples

Basic Training Examples

# 1. Kitchen environment with PTGM
python exp/trainer.py --env kitchen --scenario_type kitchenem --algorithm ptgm --lifelong append4

# 2. MMWorld with BUDS and experience replay
python exp/trainer.py --env mmworld --scenario_type mmworldem --algorithm buds --lifelong er20

# 3. Debug mode with custom epochs
python exp/trainer.py --debug --env kitchen --scenario_type debug \
    --algorithm assil --lifelong s20p10b3 --epoch 100

# 4. AsSIL with zero-shot agent
python exp/trainer.py --env kitchen --scenario_type kitchenex \
    --algorithm assil --lifelong zeros20p10b3r16g20

Advanced Configuration Examples

# 1. PTGM without t-SNE and custom clustering with larger LoRA dimension
python exp/trainer.py --env kitchen --scenario_type objective_p1 \
    --algorithm ptgm --lifelong notsnes30g40append32

# 2. IsCiL with high number of bases
python exp/trainer.py --env mmworld --scenario_type mmworldem \
    --algorithm iscil --lifelong bases200 --seed 42

# 3. Imanip with temporal replay
python exp/trainer.py --env kitchen --scenario_type kitchenex \
    --algorithm imanip --lifelong tr50 --epoch 10000

Environment-Specific Examples

Kitchen

# Object-based scenario with permutation 2
python exp/trainer.py --env kitchen --scenario_type objective_p2 \
    --algorithm ptgm --lifelong s20g20append4

MMWorld

# Easy mode with reduced epochs (n1 scenario auto-adjusts epochs)
python exp/trainer.py --env mmworld --scenario_type mmworldem \
    --algorithm buds --lifelong append16

Notes and Best Practices

1. Experiment ID: Always set a meaningful --exp_id to track your experiments
2. Seeds: Use different seeds for multiple runs to ensure reproducibility
3. Debug Mode: Use --debug for quick testing (sets epochs to 1)
4. Evaluation: Disable with --no_eval during development to save time
5. Distance Metrics:
   • maha (Mahalanobis) is default and recommended for most cases
   • euclidean and cossim force percentile thresholds

Environment Setup Requirements

Before running training, ensure environment servers are running:

# Terminal 1: Start environment server
conda activate {env_name}_eval  # e.g., kitchen_eval
python remoteEnv/{env_name}/{env_name}_server.py

# Terminal 2: Run training
conda activate silgym
python exp/trainer.py [OPTIONS]

Output Structure

Training results are saved to:

logs/{env_name}/{scenario_type}/{sync_type}/{algorithm}_{lifelong}/{date}{exp_id}seed{seed}/
├── policy/
│   └── policy_{phase}_{version}.pkl
├── skills/
│   ├── decoder_{phase}.pkl
│   └── interface_{phase}.pkl
└── experiment_config.txt