Robot platforms connected to Physical AI models

Source wording: Apptronik Apollo 2 humanoid with Inspire hands; Apptronik Apollo 2 humanoid with SharpaWave hand

Google DeepMind reports task-level success rates for Inspire-hand whole-body manipulation and SharpaWave multi-finger tasks using the same Gemini Robotics 2 checkpoint. These are developer-reported results on selected tasks, not an independent benchmark or blanket Apollo 2 capability claim.

Source wording: Unitree G1 model in MJLab simulation with 60 contact geometries per foot

Tac4Loco is trained through reinforcement learning on a Unitree G1 model in MJLab simulation, with 60 simulated contact geometries per foot. This edge records simulated embodiment coverage and must not be interpreted as training on a physical G1 or as evidence for arbitrary G1 configurations.

Source wording: Physical Unitree G1 with bilateral 60-element FSR pressure insoles

The trained actor is deployed and evaluated on one physical Unitree G1 with research-team bilateral 60-element FSR pressure insoles. Physical comparisons are source-reported, generally use ten trials per configuration, and do not establish transfer to other G1 variants, feet, sensors, payloads, or long-duration field use.

Source wording: Fourier GR-1 humanoid

The GR00T N1 paper reports real-world language-conditioned bimanual manipulation and quantitative evaluation on Fourier GR-1. It does not transfer the same results to other humanoids or later GR00T versions.

Source wording: Fourier GR-1 humanoid

The GR00T N1 paper states that pretraining uses the authors’ GR-1 humanoid data. This records training-mixture coverage only; it does not imply that every GR-1 variant, downstream checkpoint, or reported evaluation uses the same data and hardware configuration.

Source wording: Franka Emika Panda tabletop setup (fine-tuned-policy evaluation)

This relation concerns a policy fine-tuned with target demonstrations on the Franka-Tabletop setup. It is not a zero-shot result for the base OpenVLA checkpoint and does not cover every Franka installation.

Source wording: Franka Emika Panda with dual RealSense D435i cameras and two Xense Photon fingertip sensors

The Dream-Tac preprint explicitly identifies a Franka Emika Panda with two RealSense D435i cameras and two Xense Photon fingertip tactile sensors for six contact-rich tasks. The preprint does not establish independent replication or cross-robot transfer.

Source wording: Universal Robots UR5e (single and bimanual configurations)

The π0 report explicitly states that the model is trained jointly across the listed platforms, including single and bimanual UR5e configurations. This does not imply the same performance across both configurations or new UR products.

Source wording: Universal Robots UR5e (single and bimanual configurations)

The π0 report evaluates UR5e and bimanual UR5e tasks under its own pretraining and fine-tuning protocols. Results remain task-, dataset-, and configuration-specific.

Source wording: Universal Robots UR5 (represented in the Open X-Embodiment pretraining mixture)

The Octo paper’s pretraining mixture includes Berkeley Autolab UR5 data. This records training coverage for a legacy UR5 without resolving the CB controller revision and does not imply that every Octo checkpoint or upstream Open X-Embodiment snapshot uses an identical UR5 contribution.

Source wording: Universal Robots UR5 (zero-shot evaluation; not normalized to UR5e)

Octo is evaluated on two physical UR5 tasks with ten trials per task. “Zero-shot” means no target-task fine-tuning; it does not mean the UR5 embodiment was absent from pretraining, and the source does not identify this legacy UR5 as UR5e.

Source wording: Trossen WidowX 250 6-DoF (zero-shot evaluation)

The OpenVLA paper reports zero-shot evaluation on 17 BridgeData V2 WidowX tasks with 170 rollouts. BridgeData identifies the arm as a WidowX 250 6-DoF; the result remains specific to the paper protocol.

Source wording: Trossen WidowX 250 6-DoF (zero-shot evaluation)

The Octo paper lists the WidowX 250 6-DoF among its real-robot zero-shot setups. The result does not establish compatibility with other WidowX products or untested observation and action adapters.

Source wording: Google RT-series 7-DoF mobile manipulator (hardware model not disclosed)

The RT-2 paper reports about 6,000 evaluation trials on the authors’ unnamed seven-degree-of-freedom mobile manipulator. The result is tied to that platform and protocol and does not identify a commercial robot model.

Source wording: Google RT-series mobile manipulator (zero-shot evaluation)

The OpenVLA paper reports zero-shot evaluation on 12 Google-robot tasks with 60 rollouts. The hardware is normalized only to the unnamed RT-series platform family described by the source.

Source wording: Google RT-series mobile manipulator / proprietary RT-1 Robot (zero-shot evaluation)

The Octo paper identifies a proprietary RT-1 Robot as a zero-shot evaluation setup. The manufacturer and commercial product model remain undisclosed.

Source wording: Google mobile manipulator in the kitchen environment (hardware model not disclosed)

The PaLM-E paper lists the real mobile-manipulator domain in its joint embodied training mixture. The relation does not imply that PaLM-E directly emits low-level motor commands.

Source wording: Google mobile manipulator in the kitchen environment (hardware model not disclosed)

The paper evaluates embodied reasoning and long-horizon kitchen tasks on the real mobile manipulator while separate low-level policies execute generated decisions. This is not evidence of direct end-to-end motor control.

Source wording: Simulated Franka Panda with a parallel-jaw gripper and bilateral simulated GelSight Mini observations

The encoder is integrated into ACT and evaluated on eight simulated UniVTAC tasks using 100 rollouts per method-task. The author-reported 48.0% average applies to this simulated protocol and does not establish physical Panda performance, independent reproduction, or universal sensor support.

Source wording: Tianji Marvin 7-DoF arm with a parallel gripper, wrist RGB camera, and bilateral ViTai GF225 sensors

The paper reports 20 physical rollouts per method-task for Insert Tube, Insert USB, and Bottle Upright on this configuration, with human-observed binary success. The 68.3% tactile-policy average is author-reported and does not establish cross-robot transfer, independent replication, or production reliability.

Source wording: Franka Duo

Google DeepMind reports success rates for pick-and-place, tool-kitting, and insertion tasks on a Franka Duo with a Robotiq gripper. The source does not establish independent replication or performance outside the reported checkpoint, hardware configuration, and tasks.

Source wording: Fixed-base bimanual Dexmate Vega-1 with two 22-DoF Sharpa Wave hands

The T-Rex paper reports tactile-reactive midtraining on data collected with this fixed-base bimanual configuration. The edge does not imply training across multiple robot platforms, release of the complete 100-hour corpus, or transfer to another hand, sensor, or embodiment.

Source wording: Fixed-base bimanual Dexmate Vega-1 with two 22-DoF Sharpa Wave hands

The authors evaluate T-Rex on 12 contact-rich tasks with 16 randomized rollouts per task on this one research configuration. The source-reported 65% macro-average does not establish independent replication, cross-platform performance, or reliability outside the documented task and scoring protocol.

Source wording: 6-DoF RealMan RM65-B with a DH Robotics PGIA-series parallel-jaw gripper, dual 9DTact sensors, and wrist plus third-person RealSense D455 cameras

The paper reports three physical tasks with 20 binary-success trials per method-task pair on this configuration. It does not justify transfer to another RM65-B, gripper, sensor, camera layout, or task family.

Source wording: UR7e arm with an XHand five-finger tactile dexterous hand, one wrist RGB camera, and two fixed external RGB cameras

The paper supports 900 successful demonstrations collected on this platform, with 800 used in the common training pool and 100 held out for diagnostics. It does not disclose per-task counts or a public split manifest.

Source wording: UR7e arm with an XHand five-finger tactile dexterous hand, one wrist RGB camera, and two fixed external RGB cameras

The paper reports 20 rollouts per method-task under the seven-task standard protocol and 20 rollouts per method-setting under four challenges. Most reported scores are graded normalized task scores rather than binary completion.

Source wording: Franka Research 3 with a Franka Hand whose fingers are replaced by bilateral DM-Tac WS sensors, two RealSense D435i cameras, and one wrist RealSense D405

The paper reports 100 synchronized teleoperated demonstrations per task across four tasks on this one configuration. TacAura was announced but no downloadable corpus or license was verified.

Source wording: Franka Research 3 with a Franka Hand whose fingers are replaced by bilateral DM-Tac WS sensors, two RealSense D435i cameras, and one wrist RealSense D405

The paper reports 20 physical trials per model-task pair across four tasks. Baselines marked with a dagger were adapted to this tactile setup, and the fixed τ variants must not be merged into one synthetic best-task checkpoint.

Source wording: RealMan RM75B with a 3D-printed parallel gripper, bilateral DM-Tac W fingertip sensors, a wrist RealSense D405, and a first-person RealSense L515

The paper reports approximately one hour of demonstrations per subtask across eight subtasks on this configuration, but does not disclose trajectory counts, an aggregate duration, a split manifest, or a public dataset.

Source wording: RealMan RM75B with a 3D-printed parallel gripper, bilateral DM-Tac W fingertip sensors, a wrist RealSense D405, and a first-person RealSense L515

The paper reports 50 trials for each of eight clean and eight perturbed subtask settings on this configuration. It publishes per-setting results but no aggregate mean and does not establish cross-robot or cross-sensor transfer.

Source wording: Franka Emika Panda with a Robotiq 2F-140 gripper, one GelSight Mini on one gripper finger, and overhead plus wrist RealSense cameras

The paper reports task-specific RDT fine-tuning without touch and interpolant-controller training from 380 demonstrations collected through the three task pipelines on this configuration. It does not establish a tactile-native base VLA or a complete public RDT checkpoint.

Source wording: Franka Emika Panda with a Robotiq 2F-140 gripper, one GelSight Mini on one gripper finger, and overhead plus wrist RealSense cameras

The complete system is evaluated for 20 trials on each of Cup, Wipe, and Peel. The resulting 9/20, 12/20, and 7/20 end-to-end counts remain task- and hardware-specific and do not establish cross-task generalization.

Source wording: 1X humanoid

NVIDIA’s official research page says GR00T N1 demonstrates language-conditioned bimanual household manipulation on 1X humanoids. The exact 1X model and a model-specific quantitative protocol are not disclosed in the reviewed source.

Source wording: Trossen ViperX and ALOHA configurations (fine-tuned-policy evaluation)

Octo evaluates a target-data-fine-tuned policy on a ViperX setup. The ViperX embodiment was not present in Octo pretraining, so this is neither a base-checkpoint zero-shot result nor evidence that Octo was trained across ViperX.

Source wording: Trossen ViperX arms (bimanual and mobile configurations)

The π0 report says the joint training set includes bimanual and mobile setups using Trossen ViperX arms. It does not disclose an exact ViperX product code, and “based on ALOHA” does not make these setups identical to an original ALOHA system.

Source wording: Trossen ViperX and ALOHA configurations (fine-tuned-policy evaluation)

The Octo paper evaluates a physical Berkeley bimanual ALOHA setup after reinitializing a 14-dimensional action head and fine-tuning on target ALOHA demonstrations. This is not zero-shot base-checkpoint transfer, does not imply ALOHA data in Octo pretraining, and does not apply to ALOHA 2.

Source wording: Language Table setup (RT-2-PaLI-3B simulation and qualitative real-world evaluation)

The quantitative Language Table result is a simulation evaluation for RT-2-PaLI-3B. It must not be generalized to all RT-2 checkpoints or presented as a real-robot success rate.

Source wording: Language Table setup (RT-2-PaLI-3B simulation and qualitative real-world evaluation)

The paper shows qualitative real-world Language Table behavior for RT-2-PaLI-3B. The simulation metric is not transferred to this real setup.

Source wording: Language Table setup (real and simulated)

The PaLM-E training mixture includes Language Table data from simulated and real settings. This relationship records training coverage, not a claim of cross-platform deployment reliability.

Source wording: Language Table setup (real and simulated)

PaLM-E reports quantitative Language Table evaluation primarily in simulation and also presents real-tabletop behavior. The relation retains that mixed-domain boundary and does not turn simulation success into a real-robot metric.

Source wording: 23-DoF KUKA iiwa7 plus 16-DoF Allegro Hand workbench configuration with two RGB cameras (vision-only student)

ADEPT uses embodiment-specific simulation pretraining and downstream training for this KUKA-Allegro configuration. The KUKA student is vision-only, each downstream task is trained independently, and this relation does not imply physical-robot training, tactile input, or one checkpoint shared with the Flexiv-Sharpa setup.

Source wording: 23-DoF KUKA iiwa7 plus 16-DoF Allegro Hand workbench configuration with two RGB cameras (vision-only student)

The paper reports ten physical trials for each KUKA-Allegro FMB star, square-and-round, and dish condition. These author-run results are vision-only and do not establish tactile benefit, independent replication, cross-embodiment transfer, or production reliability.

Source wording: 29-DoF Flexiv Rizon plus 22-DoF Sharpa hand workbench configuration with two RGB cameras and five fingertip vision-based tactile sensors

ADEPT uses separate embodiment-specific simulation and downstream training for the Flexiv-Sharpa configuration. This relation does not imply training on a released tactile dataset, reuse of the KUKA checkpoint, or transfer to a different hand, arm, sensor, or task family.

Source wording: 29-DoF Flexiv Rizon plus 22-DoF Sharpa hand workbench configuration with two RGB cameras and five fingertip vision-based tactile sensors

In one matched square-and-round insertion condition with ten physical trials per modality, the source reports 3/10 final success for vision-only and 8/10 for visuo-tactile. The result does not establish statistical significance or generalize to other tasks, robots, hands, sensors, or operating environments.