Research & Projects

1. Sparse Reward: Sparse reward means the agent is given a distinctively high reward only upon reaching the desired final goal state, provides the possibility to obviate designing a delicate reward mechanism. It also guarantees that the agent focuses on the intended task itself without any deviation. Though born with a simple form, sparse reward diminishes the chances for policy to converge, especially in the initial random exploration stage, since the agent can hardly get positive feedbacks.

2. Hindsight: Hindsight refers to the algorithm’s ability to learn from information across goals, including ones not intended for the current task. Take the robot arm end-effector reaching task as an example. Given a goal or target position g, the end-effector may not be able to reach it under some preliminary policy. Under the sparse reward mechanism, no reward is given at the moment, and this move would therefore become a waste for training. But it's important to notice that in general, the robot is learning to reach all target position in the physical space, not just the given goal; the only purpose for the given goal is to collect valid training samples in deep reinforcement learning. Thus, we can also use the position the end-effector actually reached as the hindsight goal g' so that this sample would be valid for training, and therefore, save a significant amount of extra sample collecting.

3. Trust Region Policy Optimization: Details can be found in this paper. The optimization problem can be written as follows:

4. HTRPO Optimization Problem: Starting from the optimization problem, (1) Rewrite the problem as an expectation form w.r.b. to goals; (2) Replace the goal of the sample-collecting policy with hindsight goal g' by using importance sampling; (3) Bound the trust region with Quadratic KL-divergence; (4) Discretize the problem with Weighted Importance Sampling. Details can be found in this paper. The optimization problem can be derived into the following form, and can be solved using Karush-Kuhn-Tucker (KKT) conditions.

5. Experiment Tasks: We evaluate HTRPO in various sparse reward tasks, including simple benchmarks, image-based Atari games, and simulated robot control. To be specific, HTRPO is tested in Bit Flipping, Empty Maze, Ms. Pac-Man, Fetch Reach, Fetch Push, Fetch Slide, Fetch Pick And Place, and Sawyer Reach. Demonstrations are shown below.

6. Results: Results show HTRPO’s strong generality to different kinds of tasks and policies, and its significant improvement in performance level.

1. "Vanilla" Policy Gradient (VPG): VPG mainly adopts the characteristics of the Policy Gradient algorithm, which is introduced in the classic "REINFORCE" algorithm proposed by Richard S. Sutton. VPG here adds the word "Vanilla" to its name mainly aims to defer from other policy gradient algorithms mentioned in this project. In order to be consistent with the Actor-Critic coding frame of this project, we manually add an extra "Critic", i.e. a value function for VPG, which also coordinates with the introduction on Open AI Spinning Up. To quote from the website, the gradient of the expected finite-horizon un-discounted return of the policy is therefore written as:

The policy gradient algorithm works by updating policy parameters via stochastic gradient ascent on policy performance:

Policy gradient implementations typically compute advantage function estimates based on the infinite-horizon discounted return, despite otherwise using the finite-horizon un-discounted policy gradient formula.

2. Natural Policy Gradient (NPG): NPG is proposed in a NeurIPS 2001 paper by Sham M. Kakade. The spotlight of NPG is that it combines natural gradient with policy gradient, bounds the updates with Fisher Information Matrices, and solves the problem by taylor expansion approximation and conjugate-gradient descent. Borrowed from Katerina Fragkiadaki's slides,

3. Trust Region Policy Optimization (TRPO): TRPO is proposed a NeurIPS 2015 paper by John Schulman. It is quite similar to NPG. The main difference is the way TRPO bound the update by TV-divergence, which is then approximated by KL-divergence. It also proves that TRPO converges monotonically as the updates increase. With the solving process similar to that of NPG, the optimization problem of TRPO is:

4. Hindsight Experience Replay (HER): HER is proposed a NeurIPS 2017 paper by Marcin Andrychowicz. The concept of "Hindsight" in deep reinforcement learning was proposed for the first time. HER construct a replay buffer to collect the hindsight data from past experiences, and incorporate it with value-based and off-policy deep RL algorithms like DQN and DDPG, which can be trained with data of any offline distribution.

5. Hindsight Policy Gradient (HPG): HPG proposed by Paulo Rauber in this ICLR 2019 paper extends Hindsight from off-policy to on-policy VPG. It formalizes the policy gradient with hindsight data by leveraging goal-conditioned policy gradient and importance sampling. The hindsight policy gradient is given in the following theorem.

6. Remember and Forget for Experience Replay (ReF-ER): ReF-ER is proposed in a ICML 2019 paper by Guido Novati. ReF-ER focuses on the quality of training data in the Experience Replay buffer in general ER methods. It filters out "good" training experiences by defining "near-policy" and "far-policy" and designs concrete mechanisms for policy updating and gradient clipping.extends Hindsight from off-policy to on-policy VPG. It formalizes the policy gradient with hindsight data by leveraging goal-conditioned policy gradient and importance sampling. The hindsight policy gradient is given in the following theorem.

1. Novelty of ROI-based Grasp detection: ROI-based grasp detection (ROI-GD) considers the affiliation between the grasps and their corresponding objects in order to avoid detection error by attributing the labeled grasp to other incorrect objects, especially in object overlapping scenes. ROI-GD extract ROI features from raw object scene images and then detect grasps based on ROI features.

2. Grasp Dataset Demonstration:

3. Results Demonstration:

1. Grasp Detection: The grasp detection phase basically follow the mechanism of the project "ROI-based Robotic Grasp Detection". It detects objects and grasps on ROI features and outputs their specified affiliations to ensure a higher grasp accuracy.

2. Visual Manipulation Relationship Reasoning: Visual Manipulation Relationship regulates the order of the objects to be grasped based on their visual demonstration of spatial relationship. Normally, the ones on top of others are to be grasped first. The relationship reasoning is realized by predicting whether there are overlapping between the objects detected using the Object Pairing Pooling Layer. After figuring out the relationship between each object pair, we can construct a relationship tree describing the relationships among all objects detected in the scene.

3. Dataset Demonstration:

None at the moment...

1. Different Styles

2. Training Process

1. Phantom Module:

2. Unknown Object: