- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
The paper tackles the problem of pushing an object in clutter. separates three layers. A small fixed policy learns the inverse mapping from a local motion inten represented as a translation and rotation, plus normalized object geometry to K candidate (contact point, push direction) pairs. It is trained only on isolated simulated pushes, with no task demonstrations. They also present a planner that builds executor-aware cost-to-go fields. It runs Dijkstra over a grid with direction-dependent costs derived from coarse IK and collision probes, so routes favor motions the robot can realize. They also introduce a robot-specific executor that picks a feasible proposal and executes it.
Critical strengths
Separating a reusable local interaction policy from task-level planning and from robot-specific execution is a clean decomposition. It's nice that the same fixed policy trained on isolated pushes is used across all benchmark tasks, a second embodiment in simulation, and a real robot platform without fine-tuning. I also like the fact that inverse supervision from automatically generated simulated pushes needs no teleoperation, no task demonstrations, and no per-object dynamics identification.
Critical weaknesses
- The baselines are all run inside the authors’ own benchmark and framework, with adapters (eg, Push Anything’s reach region is replaced with the benchmark’s pusher workspace). The paper doesn’t evaluate on any benchmark the RHP or Push Anything authors defined.
- The cost fields come from a planar grid with eight directions and finite probes. This gives no feasibility guarantee, and that online validation and feedback are essential. The method also needs known object and scene geometry in advance, and it has many hand-crafted components, so the tuning burden and sensitivity are unclear.
- The setting in the real world is not practical and the scene is relatively simple (no hard clutter). The perception challenge is sidestepped as geometry is reconstructed offline and tag-based pose tracking from three cameras.
Questions for authors
- How reliable are the baselines? RHP’s long-horizon success is very low, and its failures are dominated by virtual-action validation. How would RHP and CaI compare on a benchmark and protocol defined by the RHP authors? And how much do the results depend on baseline hyperparameters that were not tuned for these maps?
- How does it scale? The policy was trained on 1,520 pushes from 36 shapes, and the held-out sets are small. How does performance change with more or fewer training pushes and shapes, and how does it fail on concave, thin, or highly elongated objects? Does a policy trained this way still work for materials or surfaces with very different friction?
Stick to the original score
Since there is no response from the authors, I will stick to my original score.