S1 Vision-Language-Action policy. Capabilities: navigate, reach on room, hallway, door, kitchen, landmark. InternVLA-N1 / DualVLN — a dual-system Vision-Language Navigation foundation model (Qwen2.5-VL-7B waypoint planner + NavDP diffusion controller). Takes egocentric RGB-D plus a natural-language navigation instruction and drives a mobile base with body-twist velocity commands, emitting STOP on arrival. Zero-shot across wheeled and legged bases (paper: Unitree Go2 / H1). Discovery view of an OpenRAL rSkill — NOT directly runnable by an agent harness; it runs via rSkill.from_pretrained + the robot HAL.
S1 Vision-Language-Action policy. Capabilities: pick, place, pick_and_place, grasp on banana, orange, apple. GR00T N1.7 (3B, Cosmos-Reason2-2B backbone) community-finetuned on the SO-101 `so101_fruit` teleop dataset (200 eps, front+wrist RGB). Emits 6-D absolute joint chunks (5 arm + gripper) over a 16-step horizon under GR00T's `new_embodiment` tag. In-process lerobot GrootPolicy, whole-model NF4 — GPU-verified 5.8 GiB peak on 8 GB. NVIDIA Open Model License (commercial OK). Discovery view of an OpenRAL rSkill — NOT directly runnable by an agent harness; it runs via rSkill.from_pretrained + the robot HAL.
S1 Vision-Language-Action policy. Capabilities: generalist, open, close, pick, place. 3D Diffuser Actor (Ke et al., 2024) — a diffusion policy over end-effector keyposes fusing multi-view RGB-D into a 3D scene representation, on the RLBench PerAct 18-task benchmark. Shares the out-of-process CoppeliaSim/PyRep sidecar with the rlbench scene backend. MIT code + checkpoints. The PerAct checkpoint is loaded verbatim; ships three live-verified starter tasks. Discovery view of an OpenRAL rSkill — NOT directly runnable by an agent harness; it runs via rSkill.from_pretrained + the robot HAL.
S1 Vision-Language-Action policy. Capabilities: insert, pick, place on peg, socket. ACT (~52M params, chunk=100) finetuned on the ALOHA bimanual sim-insertion demonstration set. Insertion is the harder ALOHA task in the original paper; the 0.20 success rate reflects that. See the norm-stats note above re: the legacy safetensors-resident buffers. Discovery view of an OpenRAL rSkill — NOT directly runnable by an agent harness; it runs via rSkill.from_pretrained + the robot HAL.
S1 Vision-Language-Action policy. Capabilities: transfer, pick, place on cube. Action Chunking Transformer (~52M-param encoder-decoder) finetuned on the ALOHA bimanual cube-transfer demonstration set. Action chunks of length 100. The mean/std norm buffers live inside model.safetensors; modern lerobot ACTPolicy drops them on load and the adapter re-applies them — see tests/sim/test_aloha_bimanual_act_aloha.py. Discovery view of an OpenRAL rSkill — NOT directly runnable by an agent harness; it runs via rSkill.from_pretrained + the robot HAL.
S1 Vision-Language-Action policy. Capabilities: pick, place, open, close on bowl, cup, drawer, object. ACT (Zhao et al., 2023) on HuggingFaceVLA/libero. ResNet-18, 4+1 enc/dec, latent VAE, chunk_size=100. Two 256x256 RGB (image / image2) + 8-D state + 7-D action; plain chunked replay. State layout matches openral_sim's LIBERO backend. Discovery view of an OpenRAL rSkill — NOT directly runnable by an agent harness; it runs via rSkill.from_pretrained + the robot HAL.
S1 Vision-Language-Action policy. Capabilities: pick, place, pick_and_place, transfer on pen. ACT (Zhao et al., 2023) finetuned for "pass the pen" pick-and-place on a real SO-101 follower arm (Apache-2.0). ResNet-18 backbone, 4+1 enc/dec, latent VAE, chunk_size=100. Emits 6-DoF absolute joint-position chunks (in joint degrees) from two RGB views (front/overview + wrist). A smaller, faster, ONNX/TensorRT-friendly sibling of rskill-smolvla-so101-pen-bf16; the whole model exports to a single ONNX graph (OPENRAL_ACT_TRT=1). Discovery view of an OpenRAL rSkill — NOT directly runnable by an agent harness; it runs via rSkill.from_pretrained + the robot HAL.
S2 decision-procedure playbook (weightless). Capabilities: plan on ambiguous reference. S2 decision procedure: when the goal is underspecified or ambiguous, resolve it from memory/scene or ask the operator a concise question before acting — never guessing on an irreversible action. Composes query_scene, memory_search, recall_object and emit_prompt. Discovery view of an OpenRAL rSkill — NOT directly runnable by an agent harness; it runs via rSkill.from_pretrained + the robot HAL.