titus-robots/titus-packing-benchmark-alpha

Titus Packing Benchmark · Alpha

A factory packing scene, robot teleoperation demonstrations, and bimanual handheld demonstrations for building and evaluating manipulation policies.

Pack rice-cake bags from a rotary table into a carton using a dual-arm YAM robot. This alpha brings the simulation environment and two existing real-world datasets together in one place, with the assets, interfaces, and documentation needed to start collaborating.

Existing SDK render illustrating the assets in a generated multi-object scene. It is not the six-bag starter reset or a policy-success result.

The collection

Open the Titus Robots collection

Component	What you get	Where to start
Simulation environment	Calibrated robot assets, rotary dish, bag and carton geometry, rigid and deformable backends, scene recipes, Python control API, RGB cameras, and a scripted single-bag reference	Download the scene archive below
Robot teleoperation	60 LeRobot v3 segments, 162,143 frames at 60 Hz; chest and two wrist cameras; bimanual absolute joint actions	rotary-table-rice-cakes
UMI demonstrations	21 retained bimanual episodes, 471 seconds; two handheld camera streams, end-effector poses, gripper widths, calibration, and MCAP recordings	trumi-box-bags-2026-10-16

The demonstration repositories are linked in place, with their existing data, licenses, and access settings. Both currently require approval through their Hugging Face dataset pages. This environment repository is publicly downloadable. Dataset guide includes pinned revisions and the important differences between the sources.

Download and run

Release alpha-v0.1.1, based on scene SDK 2026.10.28.8. The archive includes the scene assets and the pinned Linux/Python 3.12 soft-solver overlay. Isaac Sim must be installed separately.

python -m pip install -U huggingface_hub
hf download titus-robots/titus-packing-benchmark-alpha \
  --repo-type dataset --local-dir titus-packing-alpha
cd titus-packing-alpha
shasum -a 256 -c SHA256SUMS
unzip titus-packing-scene-alpha-v0.1.1.zip
cd Titus_Packing_Alpha
python3 verify_package.py

Native simulation: Linux x86-64, an NVIDIA GPU, and a pip-based Isaac Sim 6.2.0.0 environment with Python 3.12. The supplied soft backend pins Warp 1.17.0, Newton 1.7.0, MuJoCo 3.15.0, and MuJoCo-Warp 3.15.0. The lightweight SDK client runs separately. macOS can inspect assets, generate scene instances, and run CPU tests; it cannot run this native simulator package. See setup.

export ISAAC_PYTHON=/absolute/path/to/isaac-environment/bin/python
python3 -m titus setup --backend soft

# One fixed, privileged scripted task for checking your installation:
python3 -m titus reference --backend soft --out runs/reference_soft

# Build the alpha task scene without starting physics:
python3 -m titus populate recipes/alpha_six_bags_one_carton.json \
  --backend soft --seed 42 --bags 6 --boxes 1 --out runs/scene_42

# Exercise cameras and the action interface with a hold-target callback:
python3 -m examples.randomized_policy \
  --recipe recipes/alpha_six_bags_one_carton.json \
  --backend soft --seed 42 --bags 6 --boxes 1 --out runs/first_policy

Use a fresh output directory for every run. The example holds the robot's targets; replace its policy() callback to try a model. The single-bag scripted controller uses simulator state and is an installation/reference example, not a learned-policy baseline or a six-bag benchmark score.

What is inside

Path in the extracted archive	Contents
assets/	Dual-arm robot, rotary dish, workcell/carton, Sea Salt bag, textures, collision geometry, and optional factory Gaussian splat
backends/rigid/, backends/soft/	Native physics backends and their runtime locks
titus/	Scene, PolicyRunner, PlacementEvaluator, CLI, and scene-generation helpers
recipes/	Asset catalog, upstream randomized recipe, and explicit six-bag/one-carton alpha recipe
tasks/single_bag/	Scripted rigid/soft reference controllers and task-specific scorers
solver/site/	Pinned Linux/Python 3.12 dependencies with their third-party license notices
examples/, tests/, docs/	Integration examples, CPU checks, and usage guides
manifest.json, provenance/	Release file hashes and upstream source identity

Scene runtime code and supplied physical assets are preserved byte-for-byte from the SDK. This release adds documentation and a starter recipe, and excludes internal evaluation recordings and debugging handoffs. It does not include trained policies, private endpoints, provider credentials, or the separate live evaluation campaign harness.

Start with the right contract

The alpha recipe uses six bags, one carton, and no automatic carton-hold initialization. Its placement is generated from a recipe and seed. Individual generated resets have not all been validated for reachability or successful packing. The upstream randomized recipe and fixed reference have different initialization defaults; record the recipe and seed when sharing results.

PolicyRunner exposes three RGB views and robot proprioception. Scene.observe() exposes privileged simulator state for diagnostics; keep it out of ordinary policy inputs. Joint targets use radians. Gripper commands are per-finger travel in metres, not the full jaw opening; nominal jaw gap is 0.0060894 + 2 * opening_m. Physics uses 600 ticks per simulated second: 10 ticks is 60 Hz and 20 ticks is 30 Hz.

The bundled SDK's PlacementEvaluator uses assigned-carton containment, release/contact/support checks, low speed, and a two-second sampled stable dwell. The fixed single-bag reference has additional task-specific conditions. These are the supplied SDK contracts. This alpha does not ship the separate development campaign's open-top six-bag scorer, reachable-layout admission, or its experimental controller/solver variants. Results from those configurations are not directly comparable. See evaluation.

Practical tips

• Keep model and simulator environments separate. Use ISAAC_PYTHON for native simulation; do preprocessing and inference in your model process.
• Check camera order and dimensions. Simulator views are external 1024×768 and left/right wrists 256×256; teleop is three 480×270 fisheye views. An explicit preprocessing adapter is required.
• Start with one bag before scaling. Validate action units, handedness, joint order, timing, and release behavior with the reference and a small recipe.
• Record timing mode. freeze pauses simulation during inference; held_command advances under the previous command for measured inference time. Rendering and IPC are excluded by the bundled runner. Do not pool results across timing modes.
• UMI needs retargeting. Its two-camera end-effector poses are in a mapping-tag frame, not robot joints. Validate frame registration, TCP conventions, IK, and camera treatment before combining it with robot teleop.
• The factory splat is optional and visual only. It is not enabled by default and supplies no contact or support. Its alignment is estimated. Background instructions.
• Preserve physics failures. Deformable-body inversion can occur during manipulation. Keep the native guards enabled, retain the failing log/state, and classify that episode as simulation-invalid rather than reporting a policy score.

Alpha status and collaboration

This is an early environment-and-data release for integration, policy development, and reproducible experiments—not a frozen leaderboard. Known limits include deformable-contact instability; approximate simulation wrist cameras; one bag SKU and one carton asset; provisional placement geometry; and incomplete UMI-to-YAM registration. A shared scene does not establish sim-to-real fidelity. See alpha limitations and reporting.

For this packaging release, the source hashes, 13 CPU unit tests, policy-interface checks, and recipe generation were checked locally. No fresh GPU qualification was run as part of packaging; existing evaluations were left untouched. validation.json records the checks actually performed.

Useful contributions include model adapters, measured camera/robot calibration, validated UMI retargeting, reproducible contact-stability fixes, new assets with provenance, and clearly specified evaluation protocols. Open a discussion with your recipe, seed, backend, versions, timing mode, and minimal reproduction. See contributing.

Provenance and licenses

release.json records the upstream SDK identity and the pinned teleop/UMI revisions. Each existing dataset retains its MIT license. Bundled third-party software retains its own licenses; Isaac Sim is obtained separately under NVIDIA's terms. Titus scene code, documentation, and Titus-owned scene assets are released under the MIT License, matching the linked datasets. Third-party software and assets retain their applicable notices and terms; see THIRD_PARTY_NOTICES.md.