AMP1
A reference execution framework based on agentic manipulation policies. It is driven by a general-purpose AI model and requires neither low-level skills nor task-specific training, and it can run across single-arm, dual-arm and mobile platforms.
Timeline (last 90 days)
The researchers built the AMP1 reference execution framework, and the same framework runs without modification on three robot embodiments—single-arm, dual-arm and mobile manipulation—with no hardware-specific adaptation for each and no task-related fine-tuning.
The AMP1 base framework provides no low-level skills for grasping, pushing or placing; all manipulation policies are developed by the agent itself.
On the RoboLab-120 benchmark, AMP1 achieves an 85.6% success rate zero-shot, while leading VLA and WAM policies reach only 49.2% and 42.9% respectively, and AMP1 uses no demonstration data or task-specific training.
In real-world tests covering twelve tasks and three robot instances, the AMP1 agent succeeded zero-shot in 104 of 108 trials.
Compared with strong VLA and WAM baselines, AMP1 reduced the failure rate in real-world tests by 20x and 15x respectively.
After adding persistent memory storage to the base framework, the agent can carry memory representations across episodes; on the RoboDojo dual-arm manipulation tasks, the memory-equipped version reduced the failure rate by 4.6x compared with zero-shot deployment.
The researchers observed that the AMP1 agent learned to find hidden objects, use objects as tools, and calibrate the camera on its own—behaviors that were not pre-programmed but emerged as policies of the general model during interaction.
The AMPs route faces issues such as inference latency, safety and interpretability that require redesign, and AMP1 as a reference framework demonstrates feasibility but is still some distance from productization.
All Facts (8)
The researchers built the AMP1 reference execution framework, and the same framework runs without modification on three robot embodiments—single-arm, dual-arm and mobile manipulation—with no hardware-specific adaptation for each and no task-related fine-tuning.
50%UnverifiedThe AMP1 base framework provides no low-level skills for grasping, pushing or placing; all manipulation policies are developed by the agent itself.
50%UnverifiedOn the RoboLab-120 benchmark, AMP1 achieves an 85.6% success rate zero-shot, while leading VLA and WAM policies reach only 49.2% and 42.9% respectively, and AMP1 uses no demonstration data or task-specific training.
50%UnverifiedIn real-world tests covering twelve tasks and three robot instances, the AMP1 agent succeeded zero-shot in 104 of 108 trials.
50%UnverifiedCompared with strong VLA and WAM baselines, AMP1 reduced the failure rate in real-world tests by 20x and 15x respectively.
50%UnverifiedAfter adding persistent memory storage to the base framework, the agent can carry memory representations across episodes; on the RoboDojo dual-arm manipulation tasks, the memory-equipped version reduced the failure rate by 4.6x compared with zero-shot deployment.
50%UnverifiedThe researchers observed that the AMP1 agent learned to find hidden objects, use objects as tools, and calibrate the camera on its own—behaviors that were not pre-programmed but emerged as policies of the general model during interaction.
50%UnverifiedThe AMPs route faces issues such as inference latency, safety and interpretability that require redesign, and AMP1 as a reference framework demonstrates feasibility but is still some distance from productization.
50%