AI Daily ·
New Challenges in Robot Learning and AI Decision Safety
今日 AI 重點
Today's AI highlights focus on key breakthroughs and vulnerabilities in robotics and decision models. Research reveals that policy gradients in embodied RL can leak private trajectories, and natural contexts can easily flip AI decisions. On the positive side, the RAPID framework enables automated robot programming from visual demonstrations, while coding agents outperform hand-engineered planners in task and motion planning.
- 01arXivAI Safety
TRACE: Reconstructing Private Robot Trajectories from Policy Gradients in Embodied RL
While federated embodied RL assumes privacy by keeping raw sensor data local and uploading only policy gradients, this paper introduces TRACE (Temporal Reconstruction Attack on Consecutive Encodings). TRACE exploits cross-time correlations and policy-head gradient structures to autoregressively reconstruct sequential observations and actions. It achieves 18.8 dB PSNR with near-perfect action recovery, taking just 3-4.5 ms per frame. This reveals a critical vulnerability in sequential gradient sharing, calling for sequence-aware defenses.
- 02arXivRobotics
RAPID: Robot Agentic Programming from Demonstrations
Traditional robot programming is labor-intensive and lacks flexibility. RAPID bypasses this by learning from a single visual demonstration. It automatically extracts task specifications, action primitives, and a matching simulation environment. It then deploys an agentic loop to write, test, and iteratively refine control programs. By representing movements through object-centric geometric constraints rather than hardcoded paths, RAPID successfully generalized to various shapes, materials, and poses across simulations and a physical Franka robot arm.
- 03arXivRobotics
Coding Agents Outperform Hand-Engineered Planners in Task and Motion Planning
Task and Motion Planning (TAMP) is traditionally limited by hand-engineered rules. This study evaluates Claude Code (Opus 5) and Codex (GPT-5.6 Sol / GPT-6 Astra) on synthesizing generalized TAMP programs across 28 simulator environments. By interacting with simulations to calibrate strategies, the agents generated code evaluated across 98,000 episodes. The resulting programs achieved 56% to 95% success rates—outperforming classical planners (47%)—and maintained high efficiency as object scales increased, reducing computational cost by an order of magnitude.
- 04arXivAI Research
PoEM: Predicting RL Outcomes Without Re-training Existing Policies
RL post-training is computationally expensive and unstable. The PoEM framework solves this by predicting new RL policy outcomes without running any RL. The authors demonstrate that if a new reward is a linear combination of existing ones, the new log-policy is also a linear combination of the existing log-policies. Even in non-linear cases, they span a low-rank subspace. PoEM estimates weighting coefficients using only model outputs on samples, successfully synthesizing target policies across text and image tasks without any new training.
- 05arXivRobotics
AD-WM: Action-Discriminative World Models for Counterfactual MPC
Traditional world models are optimized for factual state transitions, but Model Predictive Control (MPC) requires comparing counterfactual (alternative) actions from a single state. AD-WM addresses this by combining residual latent dynamics with predictor-level action-recovery regularization based on conditional mutual information. It forces planning transitions to preserve action-dependent differences. On OGBench-Cube, AD-WM boosted hard-start success from 3.7% to 52.0%, and achieved physical zero-shot transfer on a Franka arm, raising pick-and-place success from 42.2% to 71.1% without adaptation.
- 06arXivRobotics
Rolling-WAM: Accelerating Robotic World Action Models via Rolling Imagination
Traditional World Action Models (WAMs) couple action generation with future visual prediction, but denoising the entire horizon at every step creates major latency. Rolling-WAM introduces "Rolling Imagination," maintaining a sliding window of video-action chunks at staggered noise levels. Instead of denoising from scratch, it fully denoises only the imminent action chunk for execution, while partially refining farther-future chunks. As the window slides with new observation inputs, the future chunks undergo further denoising. Tested on LIBERO, RoboTwin, and a physical Unitree G1 humanoid robot, it delivers competitive manipulation performance with a 4.5x steady-state replanning speedup.
- 07arXivAI Safety
The Invisible Trap: How Natural Context Can Easily Flip AI Decision Models
Dedicated decision models route requests and select tools, but they are highly fragile to natural context. This study introduces JevOut, demonstrating that inserting short, context-appropriate phrases—while keeping the original question and gold answer unchanged—can easily flip correct decisions. In evaluations, an optimizer successfully redirected the Jev model on 61.4% of initially correct cases, often with over 0.7 confidence in the wrong choice. Three other systems showed similar targeted flip rates (64.9%-73.2%), raising serious concerns about relying on these models for downstream actions.