Dibothrosuchus Model
A crocodylomorph-inspired quadrupedal simulation model with erect, gracile limbs and a snout-contact hunting task. Its stage 3 task uses contact from a fixed snout geom as a snap proxy; the model has no articulated jaw. Its manifest (configs/dibothrosuchus/stages.toml) declares three published behaviors on one trunk: stance (stand), locomotion (walk, warm-started from stance) and behavior (hunt — the snap task, warm-started from locomotion). See Behavior Recipes.
Generated specifications
These values are generated from the environment and compiled MJCF model. “Dynamic mass” is a simulator property, not an anatomical weight estimate.
| Gait | Quadrupedal |
|---|---|
| Observation dimension | 80 |
| Action dimension / actuators | 27 |
| Generalized coordinates (nq) | 35 |
| Generalized velocities (nv) | 34 |
| Compiled dynamic model mass | 8.7 kg |
| Plant contract revisions | Policy r7 · physics r1 · visual r1 |
| Model | environments/dibothrosuchus/assets/dibothrosuchus.xml |
Current curriculum configuration
The Stable-Baselines3 budgets and early-advancement gates below come from the current TOML files and can differ from older published runs. A stage also ends when its configured budget is exhausted. The recipe column names the behavior a stage belongs to and whether its checkpoint is a published deliverable; the warm-start column names the stage it starts from.
| Stage | Recipe | Warm-start from | Objective | SB3 configured budget | SB3 early-advancement gate |
|---|---|---|---|---|---|
| 1 — Balance | stand (deliverable) | — | Hold the erect quadrupedal stance without collapsing into a sprawl | 6M | reward ≥ 1,560; episode length ≥ 950; ≥ 10 episodes/evaluation; 3 consecutive passes |
| 2 — Locomotion | walk (deliverable) | 1 — Balance | Learn a coordinated erect-limbed diagonal-pair walk | 12M | reward ≥ 100; episode length ≥ 750; avg. velocity ≥ 0.9 m/s; ≥ 10 episodes/evaluation; 3 consecutive passes |
| 3 — Snap | hunt (deliverable) | 2 — Locomotion | Close on small prey and touch it with the snout snap proxy | 8M | reward ≥ 100; task success ≥ 50.0%; ≥ 10 episodes/evaluation; 3 consecutive passes |
Backend-specific task-success semantics
- Stable-Baselines3 — Snout-contact snap proxy: The snout snap geom contacts the prey geom while the snap reward is enabled; the model has no articulated jaw.
- JAX/MJX — Snout-tip proximity snap proxy: The snout-tip site comes within 0.12 m of the prey target position while the snap bonus is enabled; physical geom contact is not required and the model has no articulated jaw.
Per-deliverable success semantics
- stand (1 — Balance) · Stable-Baselines3 — Reward-gated stance (reward_and_length/v1): The stance checkpoint clears the reward_and_length/v1 gate: mean evaluation reward at or above the statue-derived collapse rail and a near-full-horizon mean episode length over the required consecutive evaluations. The zero-action statue clears this gate, so stand is labelled by its gate kind here rather than claimed as certified stance quality; stance_quality/v1 waits on a stance-quality and perturbation preflight this plant has not had (plan §4.8).
- walk (2 — Locomotion) · Stable-Baselines3 — Gated forward velocity (reward_and_length/v1): The locomotion checkpoint clears the reward_and_length/v1 gate: mean forward velocity at or above the stage's configured minimum, with its reward and episode-length floors, over the required consecutive evaluations.
- hunt (3 — Snap) · Stable-Baselines3 — Snout-contact snap proxy: The snout snap geom contacts the prey geom while the snap reward is enabled; the model has no articulated jaw.
- hunt (3 — Snap) · JAX/MJX — Snout-tip proximity snap proxy: The snout-tip site comes within 0.12 m of the prey target position while the snap bonus is enabled; physical geom contact is not required and the model has no articulated jaw.
Published run summaries
These are historical experiment records. A result is not evidence for the current model revision unless its provenance is explicitly marked current and verified.
No run summary has been published for this species.
Anatomy
The Dibothrosuchus model includes:
- Trunk — Level root frame with a nose-down sloping capsule and a low-slung belly
- Osteoderms — Two paramedian rows of dorsal scutes (6 pairs), the "two troughs" the genus is named for. They carry mass but are collision-disabled
- Neck + Skull — Long slender neck (pitch, yaw), narrow crested skull (head pitch), sagittal crest, and an elongate snout
- Snout — Fixed contact geom used as a snap proxy; there is no articulated jaw
- Legs — Four erect, parasagittal limbs with 5 joints each (hip pitch/roll, knee, ankle, toe) ending in a flat pad
- Tail — 4 segments, 3 actuated, carried clear of the ground in the high walk
Unique Characteristics
Dibothrosuchus is the first non-dinosaur in the project, and the smallest species in it:
- Erect crocodylomorph posture rather than the sprawl of a modern crocodilian. The limb chains are authored vertical, and every hinge joint is 0 at the home keyframe
- Hindlimbs longer than forelimbs — the reverse of Brachiosaurus, whose front legs are the longer pair
- 8.7 kg and 1.24 m nose-to-tail, an order of magnitude lighter than the T-Rex model
- Snout-referenced approach shaping — the 0.16 m skull rides a mobile neck, so approach distance is measured from the snout tip rather than the trunk
- The smallest stage-3 target in the project — a 0.06 m prey sphere, which makes aiming rather than approach the hard part of the behavior (hunt) node
Diagnostic Metrics (behavior node, hunt)
After each training stage the eval command reports:
| Metric | Description |
|---|---|
mean_forward_velocity | Average forward speed (m/s) |
gait_symmetry | Diagonal-pair stride symmetry ∈ [0, 1] |
stride_frequency | Step frequency (Hz) |
cost_of_transport | Energy efficiency (lower is better) |
mean_pelvis_height | Trunk height stability (m) |
mean_heading_alignment | cos θ toward prey ∈ [-1, 1] |
success_rate | Fraction of episodes where the fixed snout-contact proxy touches the prey target |
min_prey_distance | Closest snout approach to prey (m) |
Training Videos
Published videos for Dibothrosuchus Elaphros, where available. Artifact provenance is shown on each stage.
Balance
Hold the erect quadrupedal stance without collapsing into a sprawl
No published video is available for this stage.
Locomotion
Learn a coordinated erect-limbed diagonal-pair walk
No published video is available for this stage.
Snap
Close on small prey and touch it with the snout snap proxy
No published video is available for this stage.
Usage
cd environments/dibothrosuchus
# View the model
python scripts/view_model.py
# Train the stance node using its current TOML-configured budget
python scripts/train_sb3.py train --stage 1
# The advancing stages (stance, locomotion, behavior) in manifest order
python scripts/train_sb3.py curriculum --algorithm ppo
# Reuse an earlier run's certified stance and walk; train hunt in a fresh directory
python scripts/train_sb3.py curriculum --algorithm ppo \
--trunk-from logs/<earlier_run> --output-dir logs/<new_run>
See Behavior Recipes for the notebook workflow.