[{"id":"smolvla-libero","name":"SmolVLA - LIBERO","policy_type":"smolvla","family_name":"SmolVLA","status":"running","lifecycle":"active","action_dim":7,"description":"SmolVLA fine-tuned on LIBERO benchmarks. 7-dim action, 8-dim state, 2 cameras @ 256x256.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"evo1-libero","name":"Evo-1 - LIBERO","policy_type":"evo1","family_name":"Evo-1","status":"running","lifecycle":"active","action_dim":7,"description":"Evo-1 (InternVL3-1B vision-language backbone + flow-matching action head), one checkpoint trained on all four LIBERO suites. Published 96.65% average. 7-dim action, 8-dim state, 2 cameras @ 256x256.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"groot-libero-goal","name":"GR00T N1.7 - LIBERO Goal","policy_type":"groot","family_name":"GR00T N1.7","status":"sleeping","lifecycle":"active","action_dim":7,"description":"NVIDIA Isaac-GR00T N1.7 finetuned on the LIBERO Goal suite. NVIDIA-published metric: 195/200 (97.5%) on goal-conditioned tasks. Same architecture as groot-libero (spatial); different baked checkpoint. 7-dim delta action, 8-dim state, 2 cameras @ 256x256. Native control freq 20 Hz.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"deksprime-smolvla-datagen-pickfruit-11d818da74c54e68","name":"SmolVLA - Tabletop PickFruit (RI datagen)","policy_type":"smolvla","family_name":"SmolVLA","status":"running","lifecycle":"active","action_dim":12,"description":"First model produced end-to-end by the app/datagen service: scripted-recorded Tabletop PickFruit demos -> SmolVLA (eval 18/20=90%). PandaOmron 16-state/12-action, 3 cameras @ 256x256 (camera1/2/3).","trained_on_embodiment":"robocasa-pandaomron-12dof","reference_observation_available":true,"delisted":null},{"id":"smolvla-tabletop-pickcube","name":"SmolVLA - Tabletop PickCube (RI)","policy_type":"smolvla","family_name":"SmolVLA","status":"running","lifecycle":"active","action_dim":12,"description":"Trained end-to-end on the Robotic Inference platform: our Tabletop PickCube scene, scripted OSC expert (100% success, 300 episodes / 11,780 frames), SmolVLA finetune at 15k steps. PandaOmron 16-state/12-action, 3 cameras at 256x256.","trained_on_embodiment":"robocasa-pandaomron-12dof","reference_observation_available":true,"delisted":null},{"id":"groot-libero","name":"GR00T N1.7 - LIBERO Spatial","policy_type":"groot","family_name":"GR00T N1.7","status":"sleeping","lifecycle":"active","action_dim":7,"description":"NVIDIA Isaac-GR00T N1.7 (3B params, Cosmos-Reason2-2B / Qwen3-VL backbone, SigLip2 vision, T5 LLM, DiT action head) finetuned on LIBERO Spatial. NVIDIA-published metric: 195/200 (97.65%) success on the spatial suite. Loaded via the official `gr00t` package, not lerobot. 7-dim delta action, 8-dim state, 2 cameras @ 256x256.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"molmoact2-libero","name":"MolmoAct2 - LIBERO","policy_type":"molmoact2","family_name":"MolmoAct2","status":"running","lifecycle":"active","action_dim":7,"description":"MolmoAct2 (Ai2 action-reasoning VLA: Molmo VLM + flow-matching expert), lerobot port (policy.type=molmoact2) on the lerobot-next worker family. Single checkpoint covers all 4 LIBERO suites; lerobot-reproduced ~98.25% avg. Env-facing 7-dim delta-EE action (pads to 32 internally), 8-dim state, agentview + wrist cameras, language. ~12 GiB inference — H100-class.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"groot-libero-object","name":"GR00T N1.7 - LIBERO Object","policy_type":"groot","family_name":"GR00T N1.7","status":"sleeping","lifecycle":"active","action_dim":7,"description":"NVIDIA Isaac-GR00T N1.7 finetuned on the LIBERO Object suite. NVIDIA-published metric: 197/200 (98.45%) on object generalization tasks. Same architecture as groot-libero (spatial); different baked checkpoint. 7-dim delta action, 8-dim state, 2 cameras @ 256x256. Native control freq 20 Hz.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"openvla-libero-goal","name":"OpenVLA 7B - LIBERO Goal","policy_type":"openvla","family_name":"OpenVLA","status":"running","lifecycle":"active","action_dim":7,"description":"OpenVLA 7B (Prismatic VLM: Llama 2 7B + DINOv2 + SigLIP, action-token decoder) finetuned on LIBERO Goal. Published baseline: ~79% success. 7-dim delta action (EE pose + gripper), single 256x256 RGB agentview camera input. Inference via transformers + flash-attention-2 in bf16; bench image baked from HF checkpoint openvla/openvla-7b-finetuned-libero-goal.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"smolvla-robocasa-closefridge","name":"SmolVLA - RoboCasa CloseFridge","policy_type":"smolvla","family_name":"SmolVLA","status":"running","lifecycle":"active","action_dim":12,"description":"SmolVLA finetuned on RoboCasa CloseFridge (PandaOmron mobile manipulator). 12-dim action, 16-dim state, 3 cameras @ 256x256.","trained_on_embodiment":"robocasa-pandaomron-12dof","reference_observation_available":true,"delisted":null},{"id":"pi0fast-libero","name":"π0-FAST - LIBERO","policy_type":"pi0fast","family_name":"π0-FAST","status":"sleeping","lifecycle":"active","action_dim":7,"description":"π0-FAST autoregressive variant fine-tuned on LIBERO. 7-dim action, 3 cameras (1 empty placeholder).","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"xvla-libero","name":"X-VLA - LIBERO","policy_type":"xvla","family_name":"X-VLA","status":"running","lifecycle":"active","action_dim":7,"description":"X-VLA 0.9B Florence-2 based VLA with soft prompts. 93% avg on LIBERO. Apache 2.0.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"smolvla-libero-ri","name":"SmolVLA - LIBERO (RI-trained)","policy_type":"smolvla","family_name":"SmolVLA","status":"running","lifecycle":"active","action_dim":7,"description":"SmolVLA finetuned in-house on LIBERO via our training pipeline (from lerobot/smolvla_base). 7-dim action, 8-dim state, 2 cameras @ 256x256.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"vla-jepa-libero","name":"VLA-JEPA - LIBERO","policy_type":"vla_jepa","family_name":"VLA-JEPA","status":"running","lifecycle":"active","action_dim":7,"description":"VLA-JEPA (Qwen3-VL-2B backbone + DiT flow-matching action head; V-JEPA2 world model is training-only). lerobot-native policy (policy.type=vla_jepa) on the lerobot-next worker family (CUDA 12.8 / torch 2.8 / lerobot main 79c6821). Single checkpoint covers all 4 LIBERO suites; published ~96.5% avg. 7-dim delta-EE action, 8-dim state, agentview + wrist 224x224 cameras, language.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"openvla-libero-object","name":"OpenVLA 7B - LIBERO Object","policy_type":"openvla","family_name":"OpenVLA","status":"running","lifecycle":"active","action_dim":7,"description":"OpenVLA 7B (Prismatic VLM: Llama 2 7B + DINOv2 + SigLIP, action-token decoder) finetuned on LIBERO Object. Published baseline: ~88% success. 7-dim delta action (EE pose + gripper), single 256x256 RGB agentview camera input. Inference via transformers + flash-attention-2 in bf16; bench image baked from HF checkpoint openvla/openvla-7b-finetuned-libero-object.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"xvla-widowx","name":"X-VLA - WidowX (SIMPLER Bridge)","policy_type":"xvla","family_name":"X-VLA","status":"sleeping","lifecycle":"active","action_dim":7,"description":"X-VLA trained on Bridge (WidowX 250S). Apache-2.0. One third-person camera and an 8-dim proprio in, a 30-step chunk out, mapped by the serving adapter to the env's 7-dim delta end-effector action. The only served model that fits SIMPLER Bridge; no SIMPLER bundle runs it until P12-T7.","trained_on_embodiment":"widowx-bridge","reference_observation_available":true,"delisted":null},{"id":"groot-libero-long","name":"GR00T N1.7 - LIBERO Long (10)","policy_type":"groot","family_name":"GR00T N1.7","status":"sleeping","lifecycle":"active","action_dim":7,"description":"NVIDIA Isaac-GR00T N1.7 finetuned on the LIBERO 10 (Long) suite — long-horizon multi-step tasks. NVIDIA-published metric: 189/200 (94.35%). Same architecture as groot-libero (spatial); different baked checkpoint. 7-dim delta action, 8-dim state, 2 cameras @ 256x256. Native control freq 20 Hz.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"openvla-libero-long","name":"OpenVLA 7B - LIBERO 10 (Long-Horizon)","policy_type":"openvla","family_name":"OpenVLA","status":"running","lifecycle":"active","action_dim":7,"description":"OpenVLA 7B (Prismatic VLM: Llama 2 7B + DINOv2 + SigLIP, action-token decoder) finetuned on LIBERO 10 (Long-Horizon). Published baseline: ~53% success. 7-dim delta action (EE pose + gripper), single 256x256 RGB agentview camera input. Inference via transformers + flash-attention-2 in bf16; bench image baked from HF checkpoint openvla/openvla-7b-finetuned-libero-10.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"openvla-libero-spatial","name":"OpenVLA 7B - LIBERO Spatial","policy_type":"openvla","family_name":"OpenVLA","status":"running","lifecycle":"active","action_dim":7,"description":"OpenVLA 7B (Prismatic VLM: Llama 2 7B + DINOv2 + SigLIP, action-token decoder) finetuned on LIBERO Spatial. Published baseline: ~85% success. 7-dim delta action (EE pose + gripper), single 256x256 RGB agentview camera input. Inference via transformers + flash-attention-2 in bf16; bench image baked from HF checkpoint openvla/openvla-7b-finetuned-libero-spatial.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"fastwam-libero","name":"FastWAM - LIBERO","policy_type":"fastwam","family_name":"FastWAM","status":"running","lifecycle":"active","action_dim":7,"description":"FastWAM: a Wan2.2 video world model with a diffusion action expert. It is trained with video co-training but predicts actions directly at test time, with no future imagination. One checkpoint for all four LIBERO suites, published 96.4% average. 7-dim action, 8-dim state, 2 cameras @ 224x224. Needs an 80 GB GPU.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"pi0-libero-finetuned","name":"π0 - LIBERO (Fine-tuned)","policy_type":"pi0","family_name":"π0","status":"sleeping","lifecycle":"active","action_dim":7,"description":"π0 further fine-tuned on LIBERO for higher performance. 7-dim action, 8-dim state.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"lingbot-va-libero-long","name":"LingBot-VA - LIBERO-Long","policy_type":"lingbot_va","family_name":"LingBot-VA","status":"running","lifecycle":"active","action_dim":7,"description":"LingBot-VA: an autoregressive video-action world model on Wan2.2 that interleaves predicted video latents and actions and feeds observed keyframes back into its KV cache. Post-trained on LIBERO-Long only. 7-dim end-effector action, no proprioception, 2 cameras. Apache-2.0. Needs a 48 GB GPU.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"pi05-libero-finetuned","name":"π0.5 - LIBERO (Fine-tuned)","policy_type":"pi05","family_name":"π0.5","status":"sleeping","lifecycle":"active","action_dim":7,"description":"π0.5 fine-tuned on LIBERO (v044). 97.5% avg reported by LeRobot. Flow matching, PaliGemma 2B.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null},{"id":"multi-task-dit-libero","name":"Multi-task DiT - LIBERO","policy_type":"multi_task_dit","family_name":"Multi-task DiT","status":"sleeping","lifecycle":"active","action_dim":7,"description":"Multi-task Diffusion Transformer (~450M): CLIP ViT-B/16 vision and text conditioning and a diffusion transformer action head, no VLM. Trained by us from scratch on HuggingFaceVLA/libero (all four suites in one checkpoint). 7-dim end-effector action, 2 cameras, 8-dim state. Apache-2.0. Fits a T4.","trained_on_embodiment":"libero-panda-7dof","reference_observation_available":true,"delisted":null}]