rskill-rldx1-ft-rc365-nf4
RLDX-1 finetuned on the RoboCasa-365 cross-task generalization benchmark — 365 tasks across a wide scene/skill distribution, PandaMobile embodiment. Upstream paper-reported success: 31.5 % (vs the focused 24-task RoboCasa Kitchen finetune
RLDX-1-FT-ROBOCASA, which scores higher on its narrower suite).
This is a sibling of rldx1-ft-libero-nf4;
that README owns the canonical architecture, license, auto-managed
sidecar lifecycle, and NF4 quantization documentation for every member
of the RLDX-1 family. Read it first. The sections below cover only
this checkpoint's RC365-specific contract.
Run
The rldx adapter auto-spawns the sidecar with --embodiment-tag GENERAL_EMBODIMENT (the model card's own inference example) on first observation — single command:
openral sim run \
--config scenes/benchmarks/rldx1_rc365_pnp.yaml \
--rskill rskill://rskills/rldx1-ft-rc365-nf4 \
--view
Manual boot (debug / shared host): set OPENRAL_RLDX_AUTO_SPAWN=0 and run python tools/rldx_sidecar.py --model RLWRLD/RLDX-1-FT-RC365 --port 5555 --quantization nf4 --embodiment-tag GENERAL_EMBODIMENT yourself.
Action / state contract
RLDX-1-FT-RC365 targets PandaMobile (the Franka Panda on a mobile base — RoboCasa's default robot). Per processor_config.json's general_embodiment slot (which the model card's inference example selects via EmbodimentTag.GENERAL_EMBODIMENT):
Inputs:
| Modality | Keys | Per-key dim | Total |
|---|---|---|---|
| Video (T=4) | robot0_agentview_left, robot0_agentview_right, robot0_eye_in_hand |
256×256 RGB | 3 streams |
| State | eef_position_relative, eef_rotation_relative (quat), gripper_qpos, base_position, base_rotation (quat) |
3, 4, 2, 3, 4 | 16-D |
| Language | annotation.human.task_description |
string | — |
Outputs (action chunks of length 16, per-step layout):
| Key | Dim | Type |
|---|---|---|
end_effector_position |
3 | delta |
end_effector_rotation |
3 | delta (axis-angle) |
gripper_close |
1 | absolute |
base_motion |
4 | delta (dx, dy, dyaw, dz) |
control_mode |
1 | absolute |
| total | 12 |
The openral adapter concatenates these into a 12-D action vector. RoboCasa's PandaMobile BASIC composite consumes 11-D (arm_osc(6) + gripper(1) + base(3) + torso(1)); openral_sim/backends/robocasa.py already trims the trailing dim (the legacy "torso" slot the model treats as a control_mode flag).
State re-slicing
The openral RoboCasa scene emits the human300_16d state layout: eef_pos(3) + eef_quat(4) + base_pos(3) + base_rot(4) + grip(2). RC365's general_embodiment expects gripper BEFORE base; the adapter re-slices accordingly (see _RC365_STATE_SLICES_FROM_HUMAN300 in openral_sim.policies.rldx):
human300 idx [0:3] → state.end_effector_position_relative (eef_pos)
human300 idx [3:7] → state.end_effector_rotation_relative (eef_quat)
human300 idx [14:16] → state.gripper_qpos (grip)
human300 idx [7:10] → state.base_position (base_pos)
human300 idx [10:14] → state.base_rotation (base_rot)
Set scene.backend_options.state_layout: human300_16d on the SimEnvironment YAML to make sure the RoboCasa scene emits the layout the adapter expects.