You can deploy this model and start earning money today!

loisliowcsiro/
cyclops_runs

Name: cyclops_runs
Availability: InStock

tensorboard

robotics

reinforcement-learning

dexterous-manipulation

mujoco

apache-2.0

Views

No views yet

Cyclops Reorient — Reward Weight Ablations

Default Reward Parameters

Index	Name	Default Weight
0	orientation_alignment	30.0
1	hand_pose	0.2
2	action_rate	-0.5
3	torque	0
4	tip_slide	-0.3
5	cage_escape	-100.0
6	finger_collision	-1.0
7	hold_escalation	11.4
8	palm_detach	0.5
9	contact_count	0.0

Success threshold: 0.2 rad (~11.5°) | Hold steps: 5 | Goal switch delay: 20 steps

Ablation Runs

Original Study (experiment/without_fingerpad_tip_collision_geom branch)

Adjusted config/cyclops_finger/rsl_rl/ppo.py: entropy_coef 0.001 → 0.01, min_std 0.2 → 0.4, init_std 0.5 → 0.6.

Run	SLURM Job	orientation_alignment	hand_pose	action_rate	tip_slide	cage_escape	finger_collision	hold_escalation	palm_detach	Status
01 — default	29274656	30.0	0.2	-0.5	-0.3	-100.0	-1.0	11.4	0.5
02 — action_rate -0.3	29275079	30.0	0.2	-0.3	-0.3	-100.0	-1.0	11.4	0.5
03 — orientation_alignment 60.0	29275102	60.0	0.2	-0.5	-0.3	-100.0	-1.0	11.4	0.5
04 — hold_escalation 20.0	29275119	30.0	0.2	-0.5	-0.3	-100.0	-1.0	20.0	0.5
05 — tip_slide -0.1	29275142	30.0	0.2	-0.5	-0.1	-100.0	-1.0	11.4	0.5
06 — cage_escape -50.0	29275352	30.0	0.2	-0.5	-0.3	-50.0	-1.0	11.4	0.5
07 — cage_escape -20.0 orientation_alignment 100.0 (continued: 29402762, 29528610)	29274710	100.0	0.2	-0.5	-0.3	-20.0	-1.0	11.4	0.5	COMPLETED (final: `model_17497.pt`)
08 — cage_escape -50.0 orientation_alignment 60.0	29275361	60.0	0.2	-0.5	-0.3	-50.0	-1.0	11.4	0.5
09 — cage_escape -10.0 orientation_alignment 100.0	29354503	100.0	0.2	-0.5	-0.3	-10.0	-1.0	11.4	0.5
10 — cage_escape -5.0 orientation_alignment 100.0	29354510	100.0	0.2	-0.5	-0.3	-5.0	-1.0	11.4	0.5
11 — cage_escape -20.0 orientation_alignment 150.0 (continued: 29555347)	29354667	150.0	0.2	-0.5	-0.3	-20.0	-1.0	11.4	0.5	TIMEOUT (last: `model_11950.pt`)
15 — cage_escape -10.0 orientation_alignment 150.0	29555350	150.0	0.2	-0.5	-0.3	-10.0	-1.0	11.4	0.5	TIMEOUT (last: `model_4500.pt`)

New Study — Fingertip Contact Target Reward (main branch)

UNDESIRED_OBJECT_CONTACT_BODIES modified to include proximal_link_(center|left|right). touch_site_fingertip changed to a sphere on the sensor surface. Removed proximal_link_(center|left|right) and distal_link_(center|left|right) from DR_CONTACT_GROUP_A_GEOMS. Added base_gripper_col to SELF_COLLISION_GEOMS.

Added Fingertip_Contact_Target_Reward. Replaced Distal_Link with Fingerpad_Sensor in TIP_Collision_Geoms.

Run	SLURM Job	orientation_alignment	hand_pose	action_rate	tip_slide	cage_escape	finger_collision	hold_escalation	palm_detach	contact_count	Status
12 — cage_escape -20.0 orientation_alignment 100.0	29374461	100.0	0.2	-0.5	-0.3	-20.0	-1.0	11.4	0.5	0.0
13 — cage_escape -10.0 orientation_alignment 100.0	29374463	100.0	0.2	-0.5	-0.3	-10.0	-1.0	11.4	0.5	0.0	learning plateaued
14 — cage_escape -10.0 orientation_alignment 100.0 contact_count 2	29374469	100.0	0.2	-0.5	-0.3	-10.0	-1.0	11.4	0.5	2.0
16 — cage_escape -20.0 orientation_alignment 150.0	29567528	150.0	0.2	-0.5	-0.3	-20.0	-1.0	11.4	0.5	0.0	FAILED (exit 127: missing `--` prefix on CLI overrides, never started)

New Study — Success Curriculum Scaling with orientation_alignment and cage_escape (experiment/curriculum_reorient branch)

orientation_alignment and cage_escape rewards now scale with success_curriculum (curriculum_min=0.3). CageEscapePenalty supports curriculum_term/curriculum_min without disrupting the cage_drop termination counter. reorient_success_curriculum now uses graduated partial credit (count_threshold=2, was 3) instead of a hard binary bar.

Run	SLURM Job	orientation_alignment	hand_pose	action_rate	torque	tip_slide	cage_escape	finger_collision	hold_escalation	palm_detach	contact_count	Status
17 — cage_escape -20.0 orientation_alignment 100.0	29580281	100.0	0.2	-0.5	0	-0.3	-20.0	-1.0	11.4	0.5	0.0	RUNNING
18 — cage_escape -20.0 orientation_alignment 150.0	29580291	150.0	0.2	-0.5	0	-0.3	-20.0	-1.0	11.4	0.5	0.0	RUNNING

TensorBoard logs and checkpoints for all runs are hosted at huggingface.co/loisliowcsiro/cyclops_runs.