ROSCon 2026 · Toronto, Canada · 22–24 September 2026
Slides, links and practice resources from the session.
| My problem | Candidates | Deciding factor |
|---|---|---|
| Software-in-the-loop stack tests | Gazebo, O3DE, Isaac Sim | Control, sensors, photorealism |
| Train a policy | MuJoCo / Playground, Genesis World, Isaac Lab | Task and compute backend |
| Large, rich scenes | O3DE, Unreal | Authoring and performance |
| Custom physics | Newton, Genesis World, MuJoCo (backends) | Required solver features |
| Policy benchmarks | LIBERO, ManiSkill task suites | Benchmark fit |
| Software-in-the-loop | Robot learning | |
|---|---|---|
| Goal | Replace hardware for development and full use-case testing | Teach a skill, usually with an isolated task |
| Performance | Scene complexity, sensor fidelity and volume, robot count | Time to train, GPU batching and parallel environments |
| APIs | ROS, sockets and drivers; distributed communication | In-memory tensors; no middleware in the rollout loop |
| Time | Simulation drives the clock for the application stack | The learning algorithm steps the simulation |
| Scaling | Headless test jobs across cloud workers; real-time factor | Vectorized environment instances on the GPU |
Jetty LTS · 2025-09
26.05.0 · 2026-05-27
6.1 · 2026-09-15
3.13.0 · 2026-09
1.6.0 · 2026-09
1.4.1 · 2026-09-12
Unreal-to-Python RL integration with vectorized and multi-agent workflows. Useful for authored training environments.
Articulated-object manipulation, including visual observations. ManiSkill adds tasks, datasets and learning infrastructure.
Tests transfer across objects, spatial layouts and task goals. Useful when policy evaluation needs a defined manipulation suite.
Manipulation with contact-rich physics, plus planning, trajectory optimization and controller design. Built for optimization rather than rollouts; gradients built in.
Robot models, sensors and controllers in an integrated environment. Useful for prototyping and repeatable controller experiments.
Detailed environments, behavior authoring and visual simulation. Matters for photorealism.
Existing projects can reuse ros2-for-unity (no longer supported by Robotec.ai).
Started from a policy trained for flat ground. Increasingly difficult stairs. Compare both policies in the same evaluation setup.
Terrain, rewards, curriculum and PPO fine-tuning. Start it first and let it train in the background.
Link to be publishedVisualise the generated terrain, measure fall rate across difficulty levels and compare baseline against fine-tuned.
Link to be publishedFall rate across all terrain difficulty levels reduced from 24% to 1–2%
Same terrain, commands, camera and duration for both policies. Compare falls, foot clearance, slip and body orientation.
