An Unexpected Robot Policy: Early Evaluations of GPT-6 Astra on RoboDojo and Beyond
Read the original on arXiv Computer Vision →The Flow has not summarised this story yet — read it at arXiv Computer Vision.
The Flow has not summarised this story yet — read it at arXiv Computer Vision.
arXiv:2610.01939v1 Announce Type: new Abstract: Vision language model (VLM) agents can control robots through visual feedback and action primitives, but repeated model invocations and redundant obser...
The paper introduces Agentic Manipulation Policies (AMPs), where a generalist AI model directly controls a robot manipulator without task‑specific training or demonstrations. AMP1, a reference harness, achieves state‑of‑the‑art zero‑shot success rates on RoboLab‑120 and CaP‑Gym, outperforming specialist vision‑language‑action and world‑action models. The agent autonomously develops strategies such as searching for hidden objects and using tools, and adding a persistent memory further reduces failure rates on bimanual tasks.
RoboSPA is a large-scale robotic manipulation dataset and benchmark designed to evaluate Vision‑Language‑Action models on fine‑grained spatial reasoning and long‑horizon procedural planning. It contains 10 task categories, 56 base tasks, and 280 variants across five difficulty levels, with 527K trajectories collected from multiple embodiments and scenes. The benchmark introduces diagnostic metrics beyond binary success, revealing that current VLA models struggle with complex spatial relations, precise execution, and memory‑intensive planning.
The report evaluates the GPT-6-Astra multimodal large language model (MLLM) on the EBench platform, comparing two agent configurations: direct control, where the agent generates waypoint commands, and harness with policy, where it can invoke post‑trained vision‑language‑action (VLA) or world action model (WAM) policies. Direct control performs competitively overall, excelling on short‑horizon tasks and showing stable generalization, but falls behind on long‑horizon, tabletop, and high‑precision contact‑rich tasks. The harness with policy approach combines the strengths of both components, achieving success rates comparable to the best post‑trained policies while reducing token cost, yet still lags behind the standalone upper bounds due to suboptimal policy invocation and execution issues.
The paper evaluates GPT‑6‑Astra’s performance across 12 robotic benchmarks, showing it achieves unprecedented results when using in‑context learning from training‑set demonstrations. It highlights the benefits and costs of a dual‑system architecture (high‑level reasoning and low‑level control) and identifies practical gaps in real‑world deployment, such as expensive demonstration collection and physical constraints. The authors propose a new benchmark prototype that derives tasks from real‑world needs, omits in‑domain demos, and measures agents against a continuously running world clock.
SO-101 Bench is a simulated benchmark for language‑conditioned spatial manipulation on the low‑cost SO‑101 robotic arm, built using an Isaac Lab digital twin of a real setup. It features four tasks across five families, 138 objects with seen/unseen splits, 916 evaluation episodes, and 3,629 simulated demonstrations for policy fine‑tuning. The benchmark’s rankings mirror real‑world performance (Spearman ρ = 1.00) and reveals that while learned policies reach about 61% success, an LLM‑driven agent that writes code and delegates grasping achieves 83.2%.