Isaac Lab Manipulation

Reproductions of canonical robotic manipulation RL baselines in NVIDIA Isaac Lab using the rsl_rl PPO implementation.

The goal is reproduction, not novelty. Trained policies should match the reference behaviour and success rates reported in Isaac Lab's official benchmarks, with no modifications to the official task configurations.

This is act 1 of a four-project preparation arc for AI-Driven Rock Reshaping Simulation and Control. Each repo is independent, but together they show the route from production-stack robot-learning baselines to material simulation, contact-rich diagnostics, and graph-based fracture control.

Act	Repo	Role
1	this repo	Standard Isaac Lab manipulation reproduction. Establishes the robot-learning baseline and provides the Lift-Cube checkpoint reused by act 3.
2	excavation-rl	Granular excavation on a self-built Warp + PPO substrate. The policy did not pass the health gate; the value is the failure diagnosis.
3	cluttered-lift	Bounded Isaac Lab contact-rich manipulation diagnostic using a rigid-body granular proxy around Lift-Cube.
4	fracturegraph-control	Graph-based fracture-control prototype for AI-Driven Rock Reshaping Simulation and Control: lattice simulator, graph dynamics surrogate, and CEM planning.

Within that arc, this project is the controlled reference point. It shows that the standard manipulation stack is being used without custom reward shaping, and its trained checkpoint becomes the zero-shot and warm-start baseline for the act-3 contact-rich diagnosis.

Why This Project Exists

This is a structured reproduction project, not a tutorial run-through. Each task is reproduced under strict discipline: no reward shaping, no custom hyperparameters, no shipping at "looks reasonable" reward curves. A task counts as done only when the reference success target is matched on the evaluation distribution.

The intent is to demonstrate three things:

1. Competence with the production-style RL stack used in current robot-learning work: Isaac Lab for scalable simulation, rsl_rl for PPO training, and seed-replicated evaluation rather than reward-curve eyeballing.
2. Precise reading of upstream code. See docs/isaac-lift-cube-franka-v0.md, "Eval methodology" section, for why the obvious "Play vs base config" distinction is mostly cosmetic on this task, and what the actual randomization surface is.
3. Honest, seed-replicated reporting. Every headline is computed from N randomized rollouts on a fresh seed, not declared from a rising training-reward curve.

Status

• Isaac-Lift-Cube-Franka-v0: trained 1500 iters in 24.2 min on A10G, eval 100% success @ 2 cm goal over 256 rollouts × 2 seeds, on the randomized training distribution. (details, wandb, mp4)

Tech Stack

• Isaac Lab / Isaac Sim
• rsl_rl PPO
• Wandb / TensorBoard logging
• A10G GPU on EC2 g5.xlarge for training

Reproduction Targets

Task	Env ID	Reference success
Franka Lift Cube	Isaac-Lift-Cube-Franka-v0	> 90% (Isaac Lab default)

Layout

isaac-lab-manipulation/
├── configs/        # any per-task config overrides (kept minimal)
├── scripts/        # launchers (train.sh, play.sh, record_video.sh)
├── docs/
│   └── session_handoffs/
└── results/
    ├── figures/    # training curves, eval plots
    └── videos/     # play-mode videos

Detailed install and repro commands per task live in docs/<task>.md.