Configuration

TyGrit uses a single TOML file to configure all subsystems. See config/grasp_example.toml for an example.

Loading

from TyGrit.config import load_config

config = load_config("config/my_config.toml")
# config.env, config.scene, config.gaze, config.grasping,
# config.segmentation, config.planner, config.subgoal,
# config.mpc, config.scheduler

Only sections present in the TOML are overridden; missing sections use defaults.

Sections

[scene] — Pointcloud scene processing

Parameter	Type	Default	Description
ground_z_threshold	float	0.3	Points below this Z are ground
depth_range	[float, float]	[0.2, 3.0]	Valid depth range (metres)
enable_ground_filter	bool	true	Remove ground points
merge_radius	float	0.03	Merge radius for scene updates
downsample_voxel_size	float	0.05	Voxel grid downsample size
crop_radius	float	2.5	Keep points within this radius

[gaze] — Head tracking controller

Parameter	Type	Default	Description
lookahead_window	int	80	Number of trajectory steps to look ahead
decay_rate	float	0.99	Temporal decay for joint priorities
velocity_weight	float	1.0	Weight for velocity-based attention

[gaze.joint_priorities]

Per-joint attention weights:

Joint	Default
base	3.0
torso	2.0
shoulder_lift	1.3
elbow	1.1
wrist_flex	1.0
gripper	1.0

[grasping] — GraspGen predictor

Parameter	Type	Default	Description
checkpoint_config_path	str	—	Path to GraspGen weights YAML
num_grasps	int	200	Number of grasp candidates to generate
topk_num_grasps	int	100	Top-K grasps to keep
min_grasps	int	40	Minimum viable grasps
max_tries	int	6	Retry attempts
remove_outliers	bool	true	Filter outlier grasps
score_threshold	float	0.0	Minimum grasp score
max_grasps	int	50	Final output cap
device	str	“cuda”	Torch device

[segmentation] — Segmentation backend

Parameter	Type	Default	Description
backend	str	“sim”	Backend: "sim" (ground-truth) or "sam3" (SAM 3)

Additional parameters when backend = "sam3":

Parameter	Type	Default	Description
model_name	str	“sam3_hiera_large”	SAM 3 model variant
text_prompt	str	“object”	Text prompt for segmentation
threshold	float	0.5	Score threshold
mask_threshold	float	0.5	Mask binarization threshold
device	str	“cuda”	Torch device

[subgoal] — Subgoal generator

Parameter	Type	Default	Description
task	str	“grasp”	Task type: "grasp"

Additional parameters when task = "grasp":

Parameter	Type	Default	Description
approach_threshold	float	0.15	Joint-space distance to consider approach complete
lift_height	float	0.15	Torso raise for lift phase
lift_threshold	float	0.15	Joint-space distance to consider lift complete

[planner] — VAMP motion planner

Parameter	Type	Default	Description
timeout	float	5.0	Planning timeout (seconds)
point_radius	float	0.03	Collision sphere radius for pointcloud obstacles

[mpc] — Model predictive controller

Parameter	Type	Default	Description
gain	float	2.5	Overall tracking gain
v_max	float	5.0	Max linear velocity
w_max	float	5.0	Max angular velocity
state_weights	[float; 11]	[20, 20, 15, 12, …]	Per-DOF state error weights
control_weights	[float; 10]	[0.5, 0.8, 1, …]	Per-DOF control effort weights
joint_vel_max	[float; 8]	[2, 7, 7, …]	Joint velocity limits (torso + 7 arm)

[scheduler] — Receding-horizon loop

Parameter	Type	Default	Description
steps_per_iteration	int	10	Simulation steps per control iteration
waypoint_lookahead	int	2	Waypoints to look ahead on the path