WorldScape RSI Hits 46.58 on RoboDojo—Beats GPT-6 Astra

International access: https://4sapi.com
Domestic access: https://4sapi.org

Abstract

Manifold AI’s WorldScape RSI, short for Recursive Robot Self‑Improvement, has set a new state‑of‑the‑art record on the RoboDojo Agent Track benchmark, scoring 46.58 and surpassing GPT‑6 Astra’s 28.97. Unlike static robot policy models that freeze after training, RSI introduces a two‑layer recursive agent loop: an inner reactive robot agent executes physical manipulation tasks, while an outer reflective recursion agent reviews execution traces, refines reusable manipulation skills, and feeds validated improvements back into the runtime policy. This framework enables robots to learn directly from failed trials without additional model fine‑tuning or manual code edits from human engineers. This article breaks down the RoboDojo evaluation standard, compares quantitative benchmark metrics between WorldScape RSI and competing models, unpacks the RSI recursive loop design, analyzes in‑context learning (ICL) capabilities from WorldScape Policy 2.0, and explains how the stacked combination of ICL and recursive self‑review unlocks strong generalization for unseen robotic manipulation tasks. All original benchmark figures, task success rates and experimental comparisons are preserved, with structured interpretation of the engineering tradeoffs for embodied AI agent systems.

1. RoboDojo: The Benchmark for Embodied Agent Intelligence

Released in July 2026, RoboDojo is a collaborative benchmark initiative led by the Multi‑Media Lab (MMLab) at Tsinghua University, alongside researchers from UC Berkeley and nearly 20 other academic institutions. Built on the RoboTwin simulation engine, RoboDojo separates its evaluation suite into 42 simulation tasks and 18 real‑robot tasks. Evaluations span five core capability dimensions for bimanual robotic platforms: fine manipulation, long‑horizon execution, memory retention, open vocabulary understanding and generalization to unseen scenarios.

This benchmark tests the full end‑to‑end agent pipeline, rather than isolated vision or action modules. A qualified embodied agent must interpret natural language instructions, build scene representations, make sequential action decisions, call supporting tool modules and execute physical motor commands autonomously. It cannot rely on pre‑scripted motion sequences or human intervention mid‑run.

The baseline gap between human operators and existing robotic systems is substantial. Human participants achieve a 76.0% success rate in simulation and a perfect 100% rate on real hardware, while the average simulation success rate for the 30 representative evaluated robotic systems sits at only 8.80%, and real‑robot average success falls to 12.8%. This 57 percentage point gap is the core challenge RoboDojo is designed to quantify, which is why the benchmark is widely recognized as the “Mount Everest” for embodied agent research.

Before the WorldScape RSI release, GPT‑6 Astra represented one of the stronger public baselines on this track, yet it only reached 28.97 overall score. Its fine manipulation success rate stood at merely 4%, long‑horizon task success at 8.25%, and 16 out of all benchmark tasks could not be completed in a single trial. By contrast, WorldScape RSI pushes overall score to 46.58. Its fine manipulation success reaches 15.5%, nearly four times GPT‑6 Astra’s figure; long‑horizon task success rises to 34.25%, also a 4x improvement. Its standout metric is generalization performance on unseen tasks, hitting 35.83%, pulling ahead of all runner‑up models by a meaningful margin.

This generalization metric is the critical signal. A capable embodied agent system should not only repeat pre‑trained manipulation trajectories. It should parse unfamiliar object layouts, retain task state memory across long sequences, and adapt hand‑eye coordination patterns to new physical setups. The recursive self‑improvement design of RSI targets exactly this requirement.

2. RSI Framework: Recursive Execution, Review and Skill Refinement

RSI is not a new standalone foundation model. It adds a reflective recursive loop around an existing robot policy stack, structured as two cooperating agent layers: the inner Robot Reactive Agent and the outer Recursion Reflective Agent.

The inner layer, Robot Reactive Agent, is the runtime execution component. It consumes multi‑view RGB camera frames, robot joint state readings, natural language task prompts and historical episode memory as context. It invokes the WorldScape Policy (WSP), a skill execution policy, to emit motor commands. After each action step, control returns to the robot agent. It observes updated scene state and selects the next operation iteratively. This loop continues until task completion, timeout or failure.

The outer Recursion Reflective Agent operates offline after each full execution attempt. It loads the complete trajectory log from the trial, diagnoses failure modes, revises the underlying skill definitions, and packages updated skills into a Harness layer that the inner agent can load in subsequent attempts. Validated trajectory records are also converted into training samples, used to fine‑tune the WorldScape Policy itself over repeated cycles. The whole workflow forms the “execute → review → verify → self‑improve” cycle.

Two canonical task cases illustrate the division of labor between layers.

The first example is the push task on RoboDojo. The robot must reposition T‑shaped objects inside a target zone, with orientation and contact constraints, capped at a budget of 600 primitive action steps. On its first attempt, the inner agent uses an incorrect grasping pose. It fails to synchronize translation and rotation, releases the object prematurely, and hits 723 steps — well above the budget, marking the trial as failed. The outer reflective agent reviews the trace and isolates the root failure: poor grasp orientation, misaligned rotation timing, and premature release. It edits the corresponding skill, splitting motion into sequenced translation and rotation phases and adding a stable hold period before release. The revised skill is pushed back to the inner agent. In the retry, the task finishes within 596 primitive steps. Further iterative optimization reduces the trajectory length down to 389 steps, cutting total action overhead by roughly 35%.

Notably, the revised skill transfers to three unseen randomized layouts of the same task. The agent completes all three variants within the action budget. The outer layer is not hardcoding a single fixed solution for one scene. It distills a reusable manipulation schema, logged as editable, inspectable skill definitions rather than opaque weight updates. The robot gains a verifiable “revision workflow” for its manipulation primitives.

The second example is the packing task, which demands deeper policy adaptation beyond skill edits. In this scenario, the reflective agent continues training the underlying WorldScape Policy after skill revisions. The inner agent runs trajectories using the updated WSP. Failed attempts are appended to the training dataset for the recursion agent, creating a closed loop where both discrete skill rules and continuous action policy weights improve together. Skills encode explicit operational logic, while the policy learns low‑level continuous motor control. This dual pathway separates declarative skill knowledge from motor‑level procedural behavior.

This separation solves a common pain point for robotic learning. Pure end‑to‑end policy training is data‑hungry and difficult to audit. Skill‑only systems struggle with continuous, high‑precision servo control. RSI combines both: explicit skill harnesses for high‑level task logic and learned policies for continuous motion generation.

3. ICL from WorldScape Policy 2.0: Foundation for Rapid Task Adaptation

The strong benchmark performance of RSI builds on prior progress from WorldScape Policy 2.0, which introduced robust in‑context learning (ICL) for robotic manipulation. Traditional robotic learning pipelines require large task‑specific datasets, repeated fine‑tuning cycles and validation runs for every new manipulation scenario, which pushes deployment latency and engineering cost high. ICL reduces this startup burden. The model learns new manipulation patterns by consuming a small set of demonstration trajectories as prompt context, without parameter updates.

Public benchmark data from WorldScape Policy 2.0 demonstrates its advantage against DreamZero and π0.5 baselines across real‑robot manipulation tasks:

The Shell Game, a memory‑dependent manipulation task, highlights the gap sharply. WorldScape Policy 2.0 reaches 75% success, compared to 30% for DreamZero and 50% for π0.5. Block stacking, which requires precise spatial alignment, shows a 3x to 3.5x improvement over the other two baselines.

Three core capabilities power this ICL stack. First is action transfer: manipulation primitives transfer across objects of different shapes, materials, weights and surface friction. The model extracts invariant motion logic rather than memorizing object‑specific trajectories. Second is long‑horizon temporal reasoning. It maintains state across multi‑stage workflows, tracking intermediate goals and partial completion progress over extended action sequences. Third is episodic memory and visual state tracking. The framework integrates global memory banks and local visual context buffers, so the robot retains information about prior observations even when objects move out of the camera view temporarily.

This ICL component acts as the first adaptation layer for new tasks. It lets the robot bootstrap a working manipulation strategy from a small number of human demonstrations. The RSI recursive loop acts as the second optimization layer, refining that initial strategy by reviewing failure traces automatically. Together, they create a two‑stage adaptation pipeline: rapid in‑context bootstrapping, followed by iterative self‑correction.

When running multiple embodied model variants and managing endpoint routing for robotic agent workloads, consistent traffic governance helps track inference consumption and isolate model backends. 4sapi operates as an API gateway that standardizes request routing, logging and rate limiting for heterogeneous LLM and vision model services, simplifying operational oversight for recursive agent pipelines.

4. The Combined Pipeline: ICL Bootstrapping plus RSI Recursive Review

ICL and RSI are complementary, not competing mechanisms. ICL answers the question: “How can the robot quickly pick up a new task after seeing a small number of demos?” RSI answers a second, harder question: “How can the robot improve that initial strategy on its own, after it tries and fails?”

The sequence works in this order:

1. A small set of human demonstration trajectories is loaded as context for ICL. The robot generates an initial skill and policy for the unseen task.
2. The inner reactive agent executes the generated plan on the real robot or simulation environment.
3. If the trial fails, the outer recursion agent parses the trajectory log, isolates the root cause, revises the skill harness and optionally adds the trace to the policy training set.
4. The revised skill and updated policy are loaded for the next trial.
5. The cycle repeats until the task completes reliably within the action budget.

This flow avoids two common failure modes. Systems that rely only on ICL often freeze after the first attempt. They cannot correct systematic biases in the initial inferred plan without new human demonstrations. Systems that rely only on offline reinforcement learning require huge labeled datasets and do not support fast bootstrapping for novel household or industrial manipulation tasks. The stacked ICL + RSI design addresses both limitations.

The benchmark results reflect this synergy. The 46.58 score on RoboDojo is not just a policy win. It proves that agentic reflective layers can substantially raise embodied performance when paired with strong in‑context robotic learning.

5. Engineering and Product Context for Manifold AI

WorldScape RSI is developed by Manifold AI, an independent research and engineering team focused on embodied agent systems. The team combines foundational model research, real‑robot data collection pipelines and simulation validation workflows. Its technical stack is built around the WorldScape series of policies, split into simulation and real‑robot variants, with separate weights optimized for each domain.

The project’s data pipeline mixes simulated trajectories, teleoperation human demonstrations and hardware rollouts. It maintains high‑quality multi‑modal datasets covering vision, proprioception and action sequences, which feed both ICL prompt banks and offline fine‑tuning for the core policy. The team’s focus is on generalization, rather than overfitting to narrow benchmark scenes. This is why the largest gains appear in the unseen‑task generalization metric.

For robotics product teams, the practical takeaway is clear. Embodied agent progress is moving beyond larger foundation models alone. The next frontier is agentic runtime scaffolding: reflective loops, editable skill layers, memory systems and in‑context learning interfaces that let robots improve autonomously after deployment. Robots no longer need every failure case pre‑programmed or pre‑trained into model weights. They can build their own corrective trajectories while operating.

The RoboDojo result is a marker for this shift. GPT‑6 Astra’s respectable 28.97 score shows what a general large model can do when directly mapped to robot action spaces. WorldScape RSI at 46.58 shows what a full agent stack can achieve when the model is wrapped inside a recursive self‑improvement harness.