ManiSkill Defaults

Standard ManiSkill benchmark tasks with PPO policies and optional CAGE enhancement. These demonstrate how ManiDreams wraps any existing RL policy with uncertainty-aware action selection.

---

Common Architecture

All three tasks share the same component structure:

Component	Implementation	Details
TSIP Backend	DRISBackend	ManiSkill vectorized env with DRIS state conversion
TSIP	SimulationBasedTSIP	Parallel rollout across DRIS copies
Cage	DRISCage	Goal-conditioned cost with variance penalty (lambda_var=0.1)
Solver	PolicySampler	Samples from PPO policy distribution
Sampler	PPOPolicyAdapter	Wraps trained PPO ActorCritic for stochastic sampling
Executor	ManiSkill VectorEnv	Standard gym vectorized environment

Two modes:

• Baseline (--num_samples 0): Deterministic policy output, no CAGE evaluation

• CAGE (--num_samples N): Sample N actions from policy distribution, evaluate via TSIP + DRISCage, select best

---

PushCube

Push a cube to a target location on a tabletop.

# Baseline
python examples/tasks/maniskill_defaults/main.py --task PushCube-v1

# CAGE mode
python examples/tasks/maniskill_defaults/main.py --task PushCube-v1 \
    --num_samples 16 --n_dris_copies 16

PickCube

Pick up a cube and move it to a goal position.

# Baseline
python examples/tasks/maniskill_defaults/main.py --task PickCube-v1

# CAGE mode with perturbation
python examples/tasks/maniskill_defaults/main.py --task PickCube-v1 \
    --num_samples 16 --n_dris_copies 8 \
    --lambda_var 0.2 \
    --pose_noise 0.05 0.05 0.0 0.0 0.0 0.1

PushT

Push a T-shaped block to match a target pose.

# Baseline
python examples/tasks/maniskill_defaults/main.py --task PushT-v0

# CAGE mode
python examples/tasks/maniskill_defaults/main.py --task PushT-v0 \
    --num_samples 8 --n_dris_copies 4

---

Key Arguments

Argument	Default	Description
--task	PushCube-v1	ManiSkill task ID
--num_samples	0	Number of candidate actions (0 = baseline)
--n_dris_copies	16	DRIS copies per evaluation environment
--action_chunk	2	Multi-step action chunking
--lambda_var	0.1	Variance penalty weight in DRISCage
--pose_noise	[0.02,0.02,0,0,0,0.15]	Pose perturbation for domain randomization
--physics_noise	[0.2,0.3]	Mass and friction perturbation
--checkpoint	auto-detected	Path to PPO model checkpoint
--num_episodes	20	Number of evaluation episodes

Checkpoints are auto-detected from examples/samplers/maniskill_defaults/ckpts/.